cctally 1.7.0 → 1.7.2

package/bin/_cctally_tui.py

@@ -0,0 +1,4487 @@
+"""TUI subsystem for cctally (live terminal dashboard).
+
+Eager I/O sibling: bin/cctally loads this at startup. Owns the entire
+``cctally tui`` user-facing surface plus the shared dataclasses used
+by both the TUI and the dashboard (Phase F #22 deferred these to this
+extraction so the dashboard's existing 5 dataclass shims can resolve
+through cctally's re-exported namespace transparently):
+
+- ``cmd_tui`` — ``cctally tui`` entry point. Lazy-imports ``rich``
+  inside the function body (CLAUDE.md TUI gotcha: keeps the rest of
+  the script zero-dep). Resolves ``--variant`` (2x2 grid vs
+  expressive hero), the sync/refresh interval pair, ``--render-once``
+  / ``--snapshot-module`` fixture path, the alternate-screen / cursor
+  / SIGWINCH lifecycle, and the main render loop.
+- ``TuiKeyReader`` — raw-mode keyboard reader. Uses
+  ``termios.tcgetattr`` / ``setcbreak`` + ``select.select`` to read
+  one keystroke at a time without blocking the render loop.
+  Re-installs the saved tty mode on ``__exit__`` even when the loop
+  raises.
+- ``_TuiSyncThread`` — periodic snapshot-rebuilder. Shared base
+  class subclassed inline by the dashboard's
+  ``_DashboardSyncThread`` (in ``bin/_cctally_dashboard.py`` via
+  ``c._TuiSyncThread`` resolution at class-definition time). Owns
+  the sync-interval cadence + the ``request_sync()`` / monotonic
+  budget loop.
+- ``_tui_handle_key`` — central keymap dispatcher. Routes single
+  keystrokes through the filter / search input mode, the modal
+  open/close lifecycle, and the global hotkeys (panel switching,
+  sort cycling, help, refresh, quit). Honors
+  CLAUDE.md "Global hotkeys need modal guard" — every global
+  binding gates on ``openModal is None``.
+- ``_tui_build_*`` snapshot builder family —
+  ``_tui_build_current_week``, ``_tui_build_forecast``,
+  ``_tui_build_trend``, ``_tui_build_weekly_history``,
+  ``_tui_build_sessions``, ``_tui_build_session_detail``,
+  ``_tui_build_percent_milestones``, ``_tui_build_snapshot`` —
+  read from SQLite + the cache DB and produce one immutable
+  ``DataSnapshot``. ``_tui_build_snapshot`` is the orchestrator;
+  the rest are per-panel builders the dashboard's sync thread
+  also calls (re-exported through cctally so the dashboard's
+  ``c.X`` resolution lands).
+- ``_tui_empty_snapshot`` — minimal placeholder ``DataSnapshot``
+  used by the dashboard at boot before the first sync lands; also
+  by the panel-level test harnesses via ``ns["_tui_empty_snapshot"]``.
+- ``_tui_panel_*`` panel renderer family —
+  ``_tui_panel_current_week``, ``_tui_panel_current_week_hero``,
+  ``_tui_panel_forecast``, ``_tui_panel_trend``,
+  ``_tui_panel_sessions``. Each takes a ``DataSnapshot`` + the
+  current ``RuntimeState`` + width/height/focus hints and returns
+  a list of rich-tagged text lines (NOT a rich.Panel — the variant
+  renderers box them).
+- ``_tui_modal_*`` modal renderer family —
+  ``_tui_modal_current_week``, ``_tui_modal_forecast``,
+  ``_tui_modal_trend``, ``_tui_modal_session``. Each rebuilds its
+  body from the latest ``DataSnapshot`` every tick (CLAUDE.md
+  "TUI v2 modal/input lifecycle" gotcha: modals are NOT frozen at
+  open time; sync continues while open).
+- ``_tui_render_variant_a`` / ``_tui_render_variant_b`` —
+  full-frame composers for the two layout variants
+  (``variant_a`` is the 2x2 grid; ``variant_b`` is the expressive
+  hero). Each owns the focused-border ribbon, the toast slot, the
+  modal/overlay positioning, and the header strip.
+- ``_tui_render_help`` — full-frame help overlay (a rich.Panel
+  bordered with the keymap legend).
+- ``_tui_render_modal`` — modal dispatcher; selects the
+  per-modal renderer by ``runtime.open_modal`` slot.
+- ``_tui_render_once`` — dev hook for the
+  ``--render-once --snapshot-module`` fixture path
+  (argparse-SUPPRESSed; powers ``bin/cctally-tui-test``). Builds
+  the console with ``record=True``, runs one frame, exports
+  text/SVG, and writes to stdout. Honors ``RUNTIME_OVERRIDES``
+  dict on the snapshot module per the spec's allow-list.
+- ``_tui_header_strip_a`` / ``_tui_footer_keys`` /
+  ``_tui_render_input_prompt`` — chrome helpers for header/footer
+  rows and the in-prompt input line (filter ``f`` / search ``/``).
+- ``_tui_render_toast`` — bottom-anchored toast notification line.
+- ``_tui_colortag`` / ``_tui_escape_tags`` / ``_tui_strip_tags`` /
+  ``_tui_tagged_box_lines`` / ``_tui_lines_to_text`` — markup
+  helpers for the in-house ``{name}…{/}`` tag grammar (avoids
+  rich's ``[…]`` syntax so panel content can embed literal
+  square brackets verbatim).
+- ``_tui_box_lines`` / ``_tui_bar_string`` / ``_tui_bar_color`` /
+  ``_tui_sparkline_inline`` / ``_tui_sparkline_big`` /
+  ``_tui_width_bucket`` — drawing primitives.
+- ``_tui_verdict_of`` / ``_tui_session_model_cls`` /
+  ``_tui_format_started`` / ``_tui_format_dur`` /
+  ``_tui_sort_sessions`` / ``_tui_next_sort_key`` /
+  ``_tui_apply_session_filter`` / ``_tui_sessions_title`` —
+  data-presentation helpers for the sessions panel.
+- ``_tui_sync_interval_type`` / ``_tui_refresh_interval_type`` —
+  argparse type validators for the two CLI interval flags.
+- ``_make_run_sync_now`` / ``_make_run_sync_now_locked`` —
+  shared snapshot-rebuilder closures consumed by BOTH the TUI
+  loop and the dashboard's ``POST /api/sync`` handler + periodic
+  thread (re-exported through cctally so the dashboard's
+  shim chain lands; the test harness patches
+  ``ns["_tui_build_snapshot"]`` to stub the rebuild).
+
+- Shared dataclasses (consumed by BOTH the TUI and the dashboard,
+  via cctally's eager re-export → the dashboard's existing 5
+  dataclass shims at ``bin/_cctally_dashboard.py:487-504``
+  continue resolving transparently through
+  ``sys.modules["cctally"].X``):
+  ``DataSnapshot``, ``RuntimeState``, ``TuiCurrentWeek``,
+  ``TuiTrendRow``, ``TuiSessionRow``, ``TuiSessionDetail``,
+  ``TuiPercentMilestone``, ``WeeklyPeriodRow``,
+  ``MonthlyPeriodRow``, ``BlocksPanelRow``, ``DailyPanelRow``.
+
+What stays in bin/cctally:
+- ``ForecastInputs``, ``ForecastOutput``, ``BudgetRow`` — the
+  forecast inputs/output/budget dataclasses. Used by ``_compute_forecast``
+  (whose definition stays in cctally alongside the forecast subcommand)
+  and by the TUI builder which constructs them via the module-level
+  callable shims below.
+- ``_compute_forecast``, ``_resolve_forecast_now``,
+  ``_fetch_current_week_snapshots``, ``_load_forecast_inputs``,
+  ``_apply_midweek_reset_override``, ``_sum_cost_for_range``,
+  ``_compute_cost_for_weekref``, ``_week_ref_has_reset_event`` —
+  forecast/cost-aggregation helpers, called from this sibling via
+  module-level shims (each resolves
+  ``sys.modules["cctally"].X`` at call time).
+- The ``Block`` / ``SubWeek`` dataclasses live in ``_lib_blocks``
+  and ``_lib_subscription_weeks`` (Phase A lib siblings); accessed
+  via cctally's re-export.
+
+§5.6 audit on this extraction's monkeypatch surface
+(``tests/test_dashboard_*.py`` + ``tests/test_tui_*.py``: 11
+distinct ``ns["X"]`` direct-dict reads on moved symbols —
+``ns["DataSnapshot"]`` (6 sites), ``ns["WeeklyPeriodRow"]`` (3),
+``ns["MonthlyPeriodRow"]`` (3), ``ns["BlocksPanelRow"]`` (3),
+``ns["DailyPanelRow"]`` (3), ``ns["TuiCurrentWeek"]`` (2),
+``ns["_tui_empty_snapshot"]`` (2), ``ns["_tui_build_snapshot"]`` (1),
+``ns["_make_run_sync_now"]`` (1), ``ns["_make_run_sync_now_locked"]`` (1),
+plus ``monkeypatch.setitem`` on ``_tui_build_snapshot`` in
+``tests/test_dashboard_api_sync_refresh.py``). Forces the **eager
+re-export** carve-out per spec §4.8 (same precedent as Phase E
+#19/#20 + Phase F #21/#22):
+
+- ``ns["X"]`` dict-key reads on dataclass / function / class
+  objects propagate via eager re-export at sibling-load time;
+  PEP 562 ``__getattr__`` does NOT fire on ``ns["X"]`` (``ns`` is
+  the module's ``__dict__``, not the module proxy).
+- ``monkeypatch.setitem(ns, "_tui_build_snapshot", mock)`` mutates
+  cctally's namespace. ``_make_run_sync_now_locked`` calls
+  ``_tui_build_snapshot`` bare-name, which resolves in THIS
+  sibling's ``__dict__`` — so the mock would not propagate.
+  Pattern matches Phase D #17/#18 + F #21/#22: cross-call from
+  one moved function to another moved function that's also a
+  monkeypatch target routes through the
+  ``sys.modules['cctally']._tui_build_snapshot`` callable shim
+  at call time, ensuring the latest binding wins.
+
+Except-clause audit (Phase F #22's P1 lesson): all ``except`` clauses
+in the moved region are stdlib classes (``Exception``, ``ValueError``,
+``ImportError``, ``FileNotFoundError``) — NO cross-module exception
+classes. The ``except sys.modules["cctally"].X:`` form used in
+``_cctally_dashboard.py`` for ``UpdateError`` is NOT required here.
+
+``rich`` import policy: ``rich`` is lazy-imported INSIDE function
+bodies (``cmd_tui``, ``_tui_build_theme``, ``_tui_render_*``,
+``_tui_panel_*``, ``_tui_modal_*``, etc.) per CLAUDE.md TUI gotcha.
+The module level intentionally carries NO ``import rich`` or
+``from rich…`` line; ``Panel`` annotations on
+``_tui_render_help`` / ``_tui_render_modal`` are pure string
+annotations (lazy resolution via ``from __future__ import
+annotations``).
+
+``_TUI_VALID_STYLE_NAMES`` / ``_TUI_THEME_KEYS`` drift assertions
+(CLAUDE.md TUI gotcha: keep style names in sync with theme) move
+intact alongside the theme builder; the module-level assert at
+load time + the function-level cross-check inside
+``_tui_build_theme`` are preserved verbatim.
+
+``RUNTIME_OVERRIDES`` allow-list (CLAUDE.md TUI gotcha: dev-only
+fixture override) is inside ``_tui_render_once``; moved with the
+rest. Same for the ``--render-once --snapshot-module`` argparse
+dev path.
+
+Spec: docs/superpowers/specs/2026-05-13-bin-cctally-split-design.md §7.2
+"""
+from __future__ import annotations
+
+import argparse
+import dataclasses
+import datetime as dt
+import io
+import json
+import math
+import os
+import re
+import signal as _signal
+import sqlite3
+import sys
+import threading
+import time
+from dataclasses import dataclass, field
+from typing import Any
+
+
+def _cctally():
+    """Resolve the current ``cctally`` module at call-time (spec §5.5)."""
+    return sys.modules["cctally"]
+
+
+# === Module-level back-ref shims for helpers that STAY in bin/cctally ======
+# Each shim resolves ``sys.modules['cctally'].X`` at CALL TIME (not bind
+# time), so monkeypatches on cctally's namespace propagate into the moved
+# code unchanged. Mirrors the precedent established in
+# ``bin/_cctally_record.py``, ``bin/_cctally_cache.py``,
+# ``bin/_cctally_db.py``, ``bin/_cctally_update.py``, and
+# ``bin/_cctally_dashboard.py``.
+def eprint(*args, **kwargs):
+    return sys.modules["cctally"].eprint(*args, **kwargs)
+
+
+def parse_iso_datetime(*args, **kwargs):
+    return sys.modules["cctally"].parse_iso_datetime(*args, **kwargs)
+
+
+def _now_utc(*args, **kwargs):
+    return sys.modules["cctally"]._now_utc(*args, **kwargs)
+
+
+def open_db(*args, **kwargs):
+    return sys.modules["cctally"].open_db(*args, **kwargs)
+
+
+def load_config(*args, **kwargs):
+    return sys.modules["cctally"].load_config(*args, **kwargs)
+
+
+def format_display_dt(*args, **kwargs):
+    return sys.modules["cctally"].format_display_dt(*args, **kwargs)
+
+
+def resolve_display_tz(*args, **kwargs):
+    return sys.modules["cctally"].resolve_display_tz(*args, **kwargs)
+
+
+def normalize_display_tz_value(*args, **kwargs):
+    return sys.modules["cctally"].normalize_display_tz_value(*args, **kwargs)
+
+
+def _resolve_display_tz_obj(*args, **kwargs):
+    return sys.modules["cctally"]._resolve_display_tz_obj(*args, **kwargs)
+
+
+def _apply_display_tz_override(*args, **kwargs):
+    return sys.modules["cctally"]._apply_display_tz_override(*args, **kwargs)
+
+
+def _apply_midweek_reset_override(*args, **kwargs):
+    return sys.modules["cctally"]._apply_midweek_reset_override(*args, **kwargs)
+
+
+def _compute_display_block(*args, **kwargs):
+    return sys.modules["cctally"]._compute_display_block(*args, **kwargs)
+
+
+def _compute_forecast(*args, **kwargs):
+    return sys.modules["cctally"]._compute_forecast(*args, **kwargs)
+
+
+def _resolve_forecast_now(*args, **kwargs):
+    return sys.modules["cctally"]._resolve_forecast_now(*args, **kwargs)
+
+
+def _fetch_current_week_snapshots(*args, **kwargs):
+    return sys.modules["cctally"]._fetch_current_week_snapshots(*args, **kwargs)
+
+
+def _load_forecast_inputs(*args, **kwargs):
+    return sys.modules["cctally"]._load_forecast_inputs(*args, **kwargs)
+
+
+def _sum_cost_for_range(*args, **kwargs):
+    return sys.modules["cctally"]._sum_cost_for_range(*args, **kwargs)
+
+
+def _compute_cost_for_weekref(*args, **kwargs):
+    return sys.modules["cctally"]._compute_cost_for_weekref(*args, **kwargs)
+
+
+def _week_ref_has_reset_event(*args, **kwargs):
+    return sys.modules["cctally"]._week_ref_has_reset_event(*args, **kwargs)
+
+
+def _freshness_label(*args, **kwargs):
+    return sys.modules["cctally"]._freshness_label(*args, **kwargs)
+
+
+def _get_oauth_usage_config(*args, **kwargs):
+    return sys.modules["cctally"]._get_oauth_usage_config(*args, **kwargs)
+
+
+def _aggregate_claude_sessions(*args, **kwargs):
+    return sys.modules["cctally"]._aggregate_claude_sessions(*args, **kwargs)
+
+
+def _aggregate_monthly(*args, **kwargs):
+    return sys.modules["cctally"]._aggregate_monthly(*args, **kwargs)
+
+
+def get_claude_session_entries(*args, **kwargs):
+    return sys.modules["cctally"].get_claude_session_entries(*args, **kwargs)
+
+
+def get_latest_usage_for_week(*args, **kwargs):
+    return sys.modules["cctally"].get_latest_usage_for_week(*args, **kwargs)
+
+
+def get_latest_cost_for_week(*args, **kwargs):
+    return sys.modules["cctally"].get_latest_cost_for_week(*args, **kwargs)
+
+
+def get_milestones_for_week(*args, **kwargs):
+    return sys.modules["cctally"].get_milestones_for_week(*args, **kwargs)
+
+
+def _canonicalize_optional_iso(*args, **kwargs):
+    return sys.modules["cctally"]._canonicalize_optional_iso(*args, **kwargs)
+
+
+def get_recent_weeks(*args, **kwargs):
+    return sys.modules["cctally"].get_recent_weeks(*args, **kwargs)
+
+
+def sync_cache(*args, **kwargs):
+    return sys.modules["cctally"].sync_cache(*args, **kwargs)
+
+
+# Forecast dataclass shims — used as bare-name constructors inside
+# ``_tui_build_forecast``. The classes themselves stay in bin/cctally
+# alongside the forecast subcommand (``cmd_forecast``); call-time
+# resolution keeps monkeypatches in sync.
+def ForecastInputs(*args, **kwargs):
+    return sys.modules["cctally"].ForecastInputs(*args, **kwargs)
+
+
+def ForecastOutput(*args, **kwargs):
+    return sys.modules["cctally"].ForecastOutput(*args, **kwargs)
+
+
+def BudgetRow(*args, **kwargs):
+    return sys.modules["cctally"].BudgetRow(*args, **kwargs)
+
+
+# Dashboard back-refs consumed by the TUI's snapshot builders.
+# These functions/classes live in bin/_cctally_dashboard.py (Phase F #22),
+# re-exported through bin/cctally so the shim resolves correctly.
+def _dashboard_build_blocks_panel(*args, **kwargs):
+    return sys.modules["cctally"]._dashboard_build_blocks_panel(*args, **kwargs)
+
+
+def _dashboard_build_daily_panel(*args, **kwargs):
+    return sys.modules["cctally"]._dashboard_build_daily_panel(*args, **kwargs)
+
+
+def _dashboard_build_monthly_periods(*args, **kwargs):
+    return sys.modules["cctally"]._dashboard_build_monthly_periods(*args, **kwargs)
+
+
+def _dashboard_build_weekly_periods(*args, **kwargs):
+    return sys.modules["cctally"]._dashboard_build_weekly_periods(*args, **kwargs)
+
+
+def _build_alerts_envelope_array(*args, **kwargs):
+    return sys.modules["cctally"]._build_alerts_envelope_array(*args, **kwargs)
+
+
+def _select_current_block_for_envelope(*args, **kwargs):
+    return sys.modules["cctally"]._select_current_block_for_envelope(*args, **kwargs)
+
+
+def _SnapshotRef(*args, **kwargs):
+    return sys.modules["cctally"]._SnapshotRef(*args, **kwargs)
+
+
+# Alerts back-refs.
+# Module-level __getattr__ — lazy-resolves cctally globals at attribute-access
+# time. PEP 562 fires on ``module.X``-shaped access from outside this module;
+# bare-name lookups in function bodies bypass it. Used here for the
+# non-callable ``_AlertsConfigError`` exception class (cross-module class
+# identity is required for any future ``except _AlertsConfigError:`` site)
+# and for ``Block`` / ``SubWeek`` dataclass type references that might land
+# in annotations.
+_LAZY_ATTRS = (
+    "_AlertsConfigError",
+    "Block",
+    "SubWeek",
+)
+
+
+def __getattr__(name):  # pylint: disable=invalid-name
+    if name in _LAZY_ATTRS:
+        return getattr(sys.modules["cctally"], name)
+    raise AttributeError(name)
+
+
+# ============================================================
+# ==== TUI ====                                              =
+# ============================================================
+# Live dashboard subcommand. Lazy rich import keeps the rest of the
+# script dependency-free. All TUI-specific code lives in this block.
+
+TUI_RICH_MISSING_MSG = (
+    "tui: this subcommand requires the 'rich' package.\n"
+    "install with: pip install rich\n"
+    "(or: pipx inject cctally rich)"
+)
+
+
+# Palette — frozen TUI color values.
+TUI_PALETTE = {
+    "term_bg":      "#0a0b0d",
+    "fg":           "#d7dce1",
+    "fg_dim":       "#7a8290",
+    "fg_faint":     "#4a5060",
+    "fg_bright":    "#f4f6f8",
+    "accent":       "#6fc5e0",
+    "accent_dim":   "#3d7c92",
+    "ok":           "#7bc47f",
+    "ok_dim":       "#3f7f4e",
+    "warn":         "#e8c76e",
+    "warn_dim":     "#8a7735",
+    "bad":          "#e07a7a",
+    "bad_dim":      "#873f3f",
+    "magenta":      "#c89acf",
+    "blue":         "#8ab0d9",
+    # Badge fg colors — dark tones for high contrast against the bg-dim
+    # swatches.
+    "badge_warn_fg": "#1b1405",
+    "badge_ok_fg":   "#0a1a0c",
+    "badge_bad_fg":  "#1b0808",
+}
+
+
+def _tui_build_theme():
+    """Build a rich.theme.Theme mapping named styles to TUI_PALETTE colors.
+
+    Named styles used by the renderer:
+      fg, dim, faint, bright, accent, accent.dim,
+      ok, warn, bad, magenta, blue,
+      badge.ok, badge.warn, badge.bad,
+      focused  (alias of accent, but semantically the "focused pane border"),
+      bar.ok, bar.warn, bar.bad, bar.accent, bar.track
+    """
+    from rich.style import Style
+    from rich.theme import Theme
+    p = TUI_PALETTE
+    styles_dict = {
+        "fg":          Style(color=p["fg"]),
+        "dim":         Style(color=p["fg_dim"]),
+        "faint":       Style(color=p["fg_faint"]),
+        "bright":      Style(color=p["fg_bright"], bold=True),
+        "accent":      Style(color=p["accent"]),
+        "accent.dim":  Style(color=p["accent_dim"]),
+        "ok":          Style(color=p["ok"]),
+        "warn":        Style(color=p["warn"]),
+        "bad":         Style(color=p["bad"]),
+        "magenta":     Style(color=p["magenta"]),
+        "blue":        Style(color=p["blue"]),
+        "badge.ok":    Style(color=p["badge_ok_fg"],  bgcolor=p["ok_dim"],   bold=True),
+        "badge.warn":  Style(color=p["badge_warn_fg"], bgcolor=p["warn_dim"], bold=True),
+        "badge.bad":   Style(color=p["badge_bad_fg"],  bgcolor=p["bad_dim"],  bold=True),
+        "focused":     Style(color=p["accent"], bold=True),
+        "bar.ok":      Style(color=p["ok"]),
+        "bar.warn":    Style(color=p["warn"]),
+        "bar.bad":     Style(color=p["bad"]),
+        "bar.accent":  Style(color=p["accent"]),
+        "bar.track":   Style(color=p["fg_faint"]),
+        # v2 additions (spec §5.2)
+        "chip":        Style(color=p["term_bg"], bgcolor=p["accent"], bold=True),
+        "match":       Style(color=p["term_bg"], bgcolor=p["warn"], bold=True),
+        "prompt":      Style(color=p["accent"], bold=True),
+        "caret":       Style(color=p["term_bg"], bgcolor=p["fg"]),
+    }
+    # Function-level cross-check: the literal style dict above must match the
+    # declarative _TUI_THEME_KEYS. Catches the case where someone edits the
+    # dict but forgets to update the module-level set (or vice versa). The
+    # module-level assert covers the validator↔keys axis; this covers the
+    # keys↔actual-theme axis. We check the pre-Theme dict rather than
+    # theme.styles because rich.Theme inherits DEFAULT_STYLES (markdown/log/
+    # progress/traceback/…) which would dilute the equality check.
+    assert frozenset(styles_dict.keys()) == _TUI_THEME_KEYS, (
+        "theme/keys drift: "
+        f"added={sorted(set(styles_dict) - _TUI_THEME_KEYS)} "
+        f"removed={sorted(_TUI_THEME_KEYS - set(styles_dict))}"
+    )
+    return Theme(styles_dict)
+
+
+# Style-name shorthand -> rich-style keyword. 'b', 'u', 'pulse' are CSS
+# shorthands from the reference HTML; map them to rich equivalents.
+_TUI_TAG_SHORTHAND = {
+    "b": "bold",
+    "u": "underline",
+    "pulse": "blink",  # terminal blink — approximates the CSS pulse animation.
+}
+
+# Theme-defined style names accepted by _tui_colortag (mirrors the keys of
+# _tui_build_theme()). Must be kept in sync with that function. Any tag part
+# not in this set and not in _TUI_TAG_SHORTHAND raises ValueError.
+_TUI_VALID_STYLE_NAMES = frozenset({
+    "fg", "dim", "faint", "bright",
+    "accent", "accent.dim",
+    "ok", "warn", "bad", "magenta", "blue",
+    "badge.ok", "badge.warn", "badge.bad",
+    "focused",
+    "bar.ok", "bar.warn", "bar.bad", "bar.accent", "bar.track",
+    # v2 additions
+    "chip", "match", "prompt", "caret",
+})
+
+# Declarative enumeration of every style key produced by _tui_build_theme().
+# Single source of truth — the theme builder and the module-level drift guard
+# both consult it, so adding a theme style means editing this set (and the
+# function's dict literal) in one place.
+_TUI_THEME_KEYS = frozenset({
+    "fg", "dim", "faint", "bright",
+    "accent", "accent.dim",
+    "ok", "warn", "bad", "magenta", "blue",
+    "badge.ok", "badge.warn", "badge.bad",
+    "focused",
+    "bar.ok", "bar.warn", "bar.bad", "bar.accent", "bar.track",
+    # v2 additions
+    "chip", "match", "prompt", "caret",
+})
+
+# Module-level drift guard (no rich required): every name recognised by the
+# validator must be provided by the theme. Fires at first import — so
+# `python3 -m py_compile` followed by any import of the script catches the
+# case where someone edits one side of the pair without the other, without
+# needing to launch the `tui` subcommand.
+assert _TUI_VALID_STYLE_NAMES <= _TUI_THEME_KEYS, (
+    "_TUI_VALID_STYLE_NAMES drift: "
+    f"{sorted(_TUI_VALID_STYLE_NAMES - _TUI_THEME_KEYS)} not in theme keys"
+)
+
+
+def _tui_colortag(source: str):
+    """Render a color-tag string to a rich.text.Text.
+
+    Grammar:
+      - "{name}...{/}" -> style 'name' over inner text
+      - "{n1.n2}...{/}" -> joined styles "n1 n2" (e.g. "{ok.b}" -> "ok bold")
+      - "{{" / "}}" -> literal "{" / "}"
+      - Styles must be defined in the theme (Task 3) OR be in
+        _TUI_TAG_SHORTHAND. Unknown style names raise ValueError.
+
+    The function returns a rich.text.Text (not a string) so the caller
+    can compose it into Layouts/Panels without double-escaping.
+    """
+    from rich.text import Text
+
+    out = Text()
+    stack: list[str] = []  # active style stack; top = innermost
+    buf: list[str] = []    # pending chars for the current style run
+
+    def _flush():
+        if not buf:
+            return
+        style = " ".join(stack) if stack else ""
+        out.append("".join(buf), style=style)
+        buf.clear()
+
+    i = 0
+    n = len(source)
+    while i < n:
+        c = source[i]
+        if c == "{" and i + 1 < n and source[i + 1] == "{":
+            buf.append("{")
+            i += 2
+            continue
+        if c == "}" and i + 1 < n and source[i + 1] == "}":
+            buf.append("}")
+            i += 2
+            continue
+        if c == "{":
+            _flush()
+            end = source.find("}", i + 1)
+            if end < 0:
+                raise ValueError(f"unterminated tag at offset {i}")
+            tag = source[i + 1:end]
+            if tag == "/":
+                if not stack:
+                    raise ValueError(f"unmatched closing tag at offset {i}")
+                stack.pop()
+            else:
+                # Tag name resolution: try longest whole-tag match, peeling
+                # trailing shorthands (.b/.u/.pulse) from the end. This supports
+                # both `{ok.b}` (split/compose) AND `{bar.ok.b}` (peel `.b`,
+                # then `bar.ok` is a valid whole theme key).
+                if tag in _TUI_VALID_STYLE_NAMES:
+                    stack.append(tag)
+                else:
+                    parts = tag.split(".")
+                    resolved: str | None = None
+                    # Try progressively shorter prefixes, peeling trailing
+                    # shorthand parts off the back. `prefix` must be a valid
+                    # whole theme key; all peeled parts must be in
+                    # `_TUI_TAG_SHORTHAND`.
+                    for k in range(len(parts) - 1, 0, -1):
+                        prefix = ".".join(parts[:k])
+                        suffix = parts[k:]
+                        if prefix in _TUI_VALID_STYLE_NAMES and all(
+                            s in _TUI_TAG_SHORTHAND for s in suffix
+                        ):
+                            resolved = prefix + " " + " ".join(
+                                _TUI_TAG_SHORTHAND[s] for s in suffix
+                            )
+                            break
+                    if resolved is None:
+                        # Fallback: split-and-compose; every part must be a
+                        # known shorthand or valid style name. This supports
+                        # {ok.b} and raises on unknown names.
+                        for p in parts:
+                            if (
+                                p not in _TUI_TAG_SHORTHAND
+                                and p not in _TUI_VALID_STYLE_NAMES
+                            ):
+                                raise ValueError(
+                                    f"unknown style name {p!r} in tag "
+                                    f"{{{tag}}} at offset {i}"
+                                )
+                        resolved = " ".join(
+                            _TUI_TAG_SHORTHAND.get(p, p) for p in parts
+                        )
+                    stack.append(resolved)
+            i = end + 1
+            continue
+        buf.append(c)
+        i += 1
+
+    _flush()
+    if stack:
+        raise ValueError(f"unclosed tags remaining: {stack}")
+    return out
+
+
+def _tui_escape_tags(s: str) -> str:
+    """Escape literal `{` and `}` so user input can be safely interpolated
+    into a colortag-formatted string without being parsed as style tags.
+
+    `_tui_colortag` treats `{name}…{/}` as style tags. Doubling `{` → `{{`
+    and `}` → `}}` is the colortag grammar's literal-brace escape and the
+    parser converts each pair back to a single brace. Apply this at the
+    render boundary on any string sourced from user input or external data.
+    """
+    if not s:
+        return s
+    return s.replace("{", "{{").replace("}", "}}")
+
+
+# Double-line box-drawing glyphs.
+_TUI_BOX = {
+    "tl": "╔", "tr": "╗", "bl": "╚", "br": "╝",
+    "h":  "═", "v":  "║",
+}
+
+
+def _tui_box_lines(
+    *,
+    width: int,
+    body: list[str],
+    title: str | None = None,
+    pin: str | None = None,
+) -> list[str]:
+    """Return a list of length-`width` strings forming a double-line box.
+
+    Each body line is padded (right) or truncated to interior width (= width-2).
+    Title goes left: ╔═ title ═══╗. Pin goes right-adjacent: ╔═ title ═ pin ═╗.
+    If both won't fit, drop pin; if title won't fit, drop title.
+
+    Callers who need colored glyphs should wrap the returned strings via
+    _tui_colortag on the outside — this function emits plain text.
+    """
+    if width < 4:
+        raise ValueError(f"box width too small: {width}")
+    H, V, TL, TR, BL, BR = (
+        _TUI_BOX["h"], _TUI_BOX["v"],
+        _TUI_BOX["tl"], _TUI_BOX["tr"],
+        _TUI_BOX["bl"], _TUI_BOX["br"],
+    )
+    interior = width - 2
+
+    # Top border assembly
+    def _top() -> str:
+        if title is None:
+            return TL + H * interior + TR
+        t_seg = f" {title} "
+        # Can we fit both title and pin?
+        if pin is not None:
+            p_seg = f" {pin} "
+            # Layout: TL + H + t_seg + H*fill + p_seg + H + TR
+            # width = 1 + 1 + len(t_seg) + fill + len(p_seg) + 1 + 1
+            # fill = width - 4 - len(t_seg) - len(p_seg)
+            fill = width - 4 - len(t_seg) - len(p_seg)
+            if fill >= 1:
+                return TL + H + t_seg + H * fill + p_seg + H + TR
+        # Pin dropped (or absent) — fit just the title.
+        # Layout: TL + H + t_seg + H*fill + TR
+        # width = 1 + 1 + len(t_seg) + fill + 1
+        fill = width - 3 - len(t_seg)
+        if fill >= 1:
+            return TL + H + t_seg + H * fill + TR
+        # Title too long — fall back to plain border.
+        return TL + H * interior + TR
+
+    top = _top()
+    bot = BL + H * interior + BR
+    body_rows = []
+    for line in body:
+        if len(line) > interior:
+            line = line[:interior - 1] + "…"
+        body_rows.append(V + line + " " * (interior - len(line)) + V)
+    return [top, *body_rows, bot]
+
+
+def _tui_bar_string(pct: float, width: int) -> str:
+    """Render a filled/empty bar as a string of `█` and `░`.
+
+    Coloring is the caller's job — wrap with _tui_colortag or Text.append(style=).
+    """
+    if width <= 0:
+        return ""
+    p = max(0.0, min(100.0, float(pct)))
+    full = round((p / 100.0) * width)
+    return "█" * full + "░" * (width - full)
+
+
+def _tui_bar_color(pct: float, *, thresholds=(70.0, 90.0)) -> str:
+    """Return the theme style name for the bar based on usage thresholds.
+
+    Default thresholds match the reference design (green <70, yellow 70-90,
+    red >=90). Returns one of: 'bar.ok', 'bar.warn', 'bar.bad'.
+    """
+    low, high = thresholds
+    if pct >= high:
+        return "bar.bad"
+    if pct >= low:
+        return "bar.warn"
+    return "bar.ok"
+
+
+_TUI_SPARK_GLYPHS = ["▁", "▂", "▃", "▄", "▅", "▆", "▇", "█"]
+
+
+def _tui_sparkline_inline(points: list[int]) -> str:
+    """Map 1..8 to `_TUI_SPARK_GLYPHS`; clamp out-of-range into the 0..7 index."""
+    if not points:
+        return ""
+    return "".join(_TUI_SPARK_GLYPHS[max(0, min(7, p - 1))] for p in points)
+
+
+def _tui_sparkline_big(points: list[int]) -> str:
+    """Render a 3-row block chart, 2 chars wide per point, space-separated.
+
+    Point values 1..8 scaled to a 0..9 height; distributed top-down across
+    three segments each taking values 0..3; height-per-segment maps to
+    {0:'  ', 1:'▂▂', 2:'▄▄', 3:'██'}.
+    """
+    if not points:
+        return "\n\n"
+    rows: list[list[str]] = [[], [], []]
+    glyph_map = ["  ", "▂▂", "▄▄", "██"]
+    for p in points:
+        h = max(1, min(8, int(p))) / 8 * 9  # 0..9
+        pieces = [0, 0, 0]
+        for i in (2, 1, 0):
+            if h >= 3:
+                pieces[i] = 3
+                h -= 3
+            elif h >= 2:
+                pieces[i] = 2
+                h = 0
+            elif h >= 1:
+                pieces[i] = 1
+                h = 0
+        for r_idx in range(3):
+            rows[r_idx].append(glyph_map[pieces[r_idx]])
+    return "\n".join(" ".join(r) for r in rows)
+
+
+def _tui_width_bucket(width: int) -> str:
+    """Pick a layout bucket from terminal width.
+
+    - >= 120: 'wide'    (full design, 120×36 as primary)
+    - 100..119: 'compact' (drops Model/Project in A sessions, 4wk trend in B)
+    - 80..99:  'narrow'  (same rules as compact + shows a narrow-warning line)
+    - < 80:    'refuse'  (error message, exit 1)
+    """
+    if width >= 120:
+        return "wide"
+    if width >= 100:
+        return "compact"
+    if width >= 80:
+        return "narrow"
+    return "refuse"
+
+
+# -------- data layer -----------------------------------------------------
+# Dataclasses produced by the sync thread and consumed by the render
+# thread. Treat DataSnapshot as immutable — the sync thread publishes a
+# new instance and the renderer swaps the reference atomically.
+
+
+@dataclass
+class TuiCurrentWeek:
+    week_start_at: dt.datetime
+    week_end_at: dt.datetime
+    used_pct: float
+    five_hour_pct: float | None
+    five_hour_resets_at: dt.datetime | None
+    spent_usd: float
+    dollars_per_percent: float | None
+    latest_snapshot_at: dt.datetime
+    # Freshness fields (Task C6). Computed by `_tui_build_current_week` via
+    # `_freshness_label` against the configured oauth_usage thresholds. Default
+    # None so fixture modules that construct `TuiCurrentWeek` directly without
+    # populating these stay backwards-compatible — the renderer treats `None`
+    # the same as `"fresh"` and hides the chip. Refs spec §3.4.
+    freshness_label: str | None = None
+    freshness_age: int | None = None
+    # Current 5h block snapshot for the dashboard envelope (spec §4.1). Snake-case
+    # dict with keys: block_start_at, seven_day_pct_at_block_start,
+    # seven_day_pct_delta_pp, crossed_seven_day_reset. Populated by
+    # `_tui_build_current_week` via `_select_current_block_for_envelope`; the
+    # default `None` keeps fixture modules that construct TuiCurrentWeek
+    # directly (without this field) backwards-compatible.
+    five_hour_block: dict | None = None
+
+
+@dataclass
+class TuiTrendRow:
+    week_label: str              # e.g. "Apr 14"
+    week_start_at: dt.datetime
+    used_pct: float | None       # None when the week has a cost snapshot
+                                 # but no usage snapshot (phantom weeks)
+    dollars_per_percent: float | None
+    delta_dpp: float | None      # vs prior week
+    spark_height: int            # 1..8 normalized
+    is_current: bool
+
+
+@dataclass
+class WeeklyPeriodRow:
+    """One subscription-week row for the dashboard's Weekly panel/modal.
+
+    `models` is a list of `{model, display, chip, cost_usd, cost_pct}`
+    dicts sorted by `cost_usd` descending. Pre-bucketed in Python so
+    the React layer never re-derives per-model coloring.
+    """
+    label: str                          # "04-23" — MM-DD of the week start
+    cost_usd: float
+    total_tokens: int
+    input_tokens: int
+    output_tokens: int
+    cache_creation_tokens: int
+    cache_read_tokens: int
+    used_pct: float | None              # from weekly_usage_snapshots overlay
+    dollar_per_pct: float | None        # cost / used_pct when used_pct > 0
+    delta_cost_pct: float | None        # (cost - prev_cost) / prev_cost
+    is_current: bool
+    models: list[dict[str, Any]]
+    week_start_at: str                  # ISO-8601 with tz, from SubWeek.start_ts
+    week_end_at: str                    # ISO-8601 with tz, from SubWeek.end_ts
+
+
+@dataclass
+class MonthlyPeriodRow:
+    """One calendar-month row for the dashboard's Monthly panel/modal."""
+    label: str                          # "YYYY-MM"
+    cost_usd: float
+    total_tokens: int
+    input_tokens: int
+    output_tokens: int
+    cache_creation_tokens: int
+    cache_read_tokens: int
+    delta_cost_pct: float | None
+    is_current: bool
+    models: list[dict[str, Any]]
+
+
+@dataclass
+class BlocksPanelRow:
+    """One row of the dashboard's Blocks panel.
+
+    Subset of the `Block` dataclass — drops token counts (panel is
+    cost-driven; tokens belong to a future modal), drops `entries_count`
+    / `is_gap` / `burn_rate` / `projection` (panel doesn't render them),
+    and pre-formats `label` server-side for the local-tz "HH:MM MMM DD"
+    display.
+    """
+    start_at: str          # ISO-8601 UTC
+    end_at: str            # ISO-8601 UTC, start_at + 5h
+    anchor: str            # 'recorded' | 'heuristic'
+    is_active: bool        # now_utc < end_at AND entries_count > 0
+    cost_usd: float
+    models: list[dict[str, Any]]   # ModelCostRow shape, sorted desc by cost
+    label: str             # "HH:MM MMM DD" in local tz, e.g. "14:00 Apr 26"
+
+
+@dataclass
+class DailyPanelRow:
+    """One row of the dashboard's Daily heatmap panel.
+
+    `intensity_bucket` is the server-computed quintile bucket (0..5) —
+    bucket 0 is reserved for zero-cost days; buckets 1..5 are quintiles
+    over non-zero days.
+
+    v2.3: Added per-day token rollup + `cache_hit_pct` so the Daily
+    detail modal can surface the same fields the CLI's `daily` command
+    shows. Defaults preserve compatibility with `_empty_dashboard_snapshot`
+    and any pre-v2.3 fixture that omits the new fields.
+    """
+    date: str              # local-tz YYYY-MM-DD
+    label: str             # "MM-DD" — pre-formatted, mirrors Weekly/Monthly idiom
+    cost_usd: float
+    is_today: bool
+    intensity_bucket: int  # 0..5
+    models: list[dict[str, Any]]   # ModelCostRow shape, sorted desc by cost
+    # ---- v2.3 additions: Daily modal token + cache rollup ----
+    input_tokens: int = 0
+    output_tokens: int = 0
+    cache_creation_tokens: int = 0
+    cache_read_tokens: int = 0
+    total_tokens: int = 0
+    cache_hit_pct: float | None = None
+
+
+@dataclass
+class TuiSessionRow:
+    started_at: dt.datetime
+    duration_minutes: float
+    model_primary: str           # first model used in the session
+    cost_usd: float
+    cache_hit_pct: float | None
+    project_label: str           # basename of project_path
+    session_id: str              # full session UUID (v2: needed for session-detail modal)
+
+
+@dataclass
+class TuiPercentMilestone:
+    """One row in the Current-Week per-percent modal (spec §4.6.1)."""
+    percent: int                           # 1..100
+    crossed_at: dt.datetime                # captured_at_utc
+    cumulative_cost_usd: float
+    marginal_cost_usd: float | None
+    five_hour_pct_at_crossing: float | None
+
+
+def _tui_build_percent_milestones(
+    conn: sqlite3.Connection,
+) -> list[TuiPercentMilestone]:
+    """Return per-percent crossings for the current week's ACTIVE
+    segment, ascending by percent.
+
+    Resolves `week_start_date` from the latest `weekly_usage_snapshots` row
+    — the same path `cmd_percent_breakdown` takes. The post-override
+    `TuiCurrentWeek.week_start_at` is NOT suitable here: after a mid-week
+    reset, `_apply_midweek_reset_override` shifts that datetime forward to
+    the reset instant, whose `.date()` no longer matches the `week_start_date`
+    under which milestones were recorded.
+
+    v1.7.2: when a `week_reset_events` row exists for the snapshot's
+    `week_end_at`, narrow to the active segment so the dashboard /
+    TUI milestone panel stays coherent with the already-credit-aware
+    header. ``active_segment = 0`` (sentinel) preserves legacy
+    behavior on un-credited weeks.
+
+    Returns [] if no usage snapshot exists, OR if the active segment
+    has no milestone rows yet (post-credit "fresh" state).
+    """
+    latest = conn.execute(
+        "SELECT week_start_date, week_end_at FROM weekly_usage_snapshots "
+        "WHERE week_end_at IS NOT NULL "
+        "ORDER BY captured_at_utc DESC, id DESC LIMIT 1"
+    ).fetchone()
+    if latest is None:
+        # Legacy fallback: a snapshot without week_end_at can still have
+        # milestones — keep the prior behavior in that path.
+        latest = conn.execute(
+            "SELECT week_start_date, NULL AS week_end_at "
+            "FROM weekly_usage_snapshots "
+            "ORDER BY captured_at_utc DESC, id DESC LIMIT 1"
+        ).fetchone()
+        if latest is None:
+            return []
+
+    # Resolve active segment via the canonical end_at.
+    active_segment = 0
+    if latest["week_end_at"]:
+        try:
+            canon_end = _canonicalize_optional_iso(
+                latest["week_end_at"], "tui.pm.cur"
+            )
+        except (AttributeError, ValueError):
+            canon_end = None
+        if canon_end:
+            seg_row = conn.execute(
+                "SELECT id FROM week_reset_events "
+                "WHERE new_week_end_at = ? "
+                "ORDER BY id DESC LIMIT 1",
+                (canon_end,),
+            ).fetchone()
+            if seg_row is not None:
+                active_segment = int(seg_row["id"])
+
+    rows = [
+        r for r in get_milestones_for_week(conn, latest["week_start_date"])
+        if int(r["reset_event_id"] or 0) == active_segment
+    ]
+    out: list[TuiPercentMilestone] = []
+    for r in rows:
+        try:
+            crossed = parse_iso_datetime(r["captured_at_utc"], "captured_at_utc")
+        except ValueError:
+            continue
+        out.append(TuiPercentMilestone(
+            percent=int(r["percent_threshold"]),
+            crossed_at=crossed,
+            cumulative_cost_usd=float(r["cumulative_cost_usd"]),
+            marginal_cost_usd=(float(r["marginal_cost_usd"])
+                               if r["marginal_cost_usd"] is not None else None),
+            five_hour_pct_at_crossing=(float(r["five_hour_percent_at_crossing"])
+                                       if r["five_hour_percent_at_crossing"] is not None
+                                       else None),
+        ))
+    return out
+
+
+@dataclass
+class DataSnapshot:
+    """All data needed to render one TUI frame. Produced by sync thread,
+    consumed by main thread. Treat as immutable."""
+    current_week: TuiCurrentWeek | None
+    forecast: Any | None          # ForecastOutput from _compute_forecast
+    trend: list[TuiTrendRow]
+    sessions: list[TuiSessionRow]
+    last_sync_at: float | None    # monotonic (time.monotonic())
+    last_sync_error: str | None
+    generated_at: dt.datetime     # wall-clock UTC for displayed timestamps
+    # ---- v2 additions (spec §4.5) ----
+    percent_milestones: list[TuiPercentMilestone] = field(default_factory=list)
+    weekly_history: list[TuiTrendRow] = field(default_factory=list)
+    # ---- v2.1 additions: dashboard Weekly / Monthly panels ----
+    weekly_periods:  list[WeeklyPeriodRow]  = field(default_factory=list)
+    monthly_periods: list[MonthlyPeriodRow] = field(default_factory=list)
+    # ---- v2.2 additions: dashboard Blocks / Daily panels ----
+    blocks_panel: list[BlocksPanelRow] = field(default_factory=list)
+    daily_panel:  list[DailyPanelRow]  = field(default_factory=list)
+    # ---- threshold-actions T5: snapshot alerts envelope array ----
+    # Populated at sync-thread snapshot-build time by
+    # `_build_alerts_envelope_array(conn)`. Single source of truth for
+    # both the dashboard panel (slices to 10) and the modal (renders all
+    # 100). Empty list when alerts feature is disabled, no rows have
+    # `alerted_at` set, or DB read fails (sub-build catches the exception
+    # and records it on `last_sync_error`). Stored as
+    # already-envelope-shaped dicts so `snapshot_to_envelope` stays a
+    # pure renderer (no DB I/O on the dashboard hot path; mirrors how
+    # `current_week.five_hour_block` is precomputed via
+    # `_select_current_block_for_envelope`).
+    alerts: list[dict] = field(default_factory=list)
+
+    @classmethod
+    def synthesize_for_marketing(cls, *, as_of_iso: str) -> "DataSnapshot":
+        """Build a deterministic DataSnapshot for README screenshot pipelines.
+
+        Used by tests/fixtures/readme/tui_snapshot.py when run via
+        `cctally tui --render-once --snapshot-module ...`. Numbers are
+        narratively coherent with the marketing fixture's stats.db /
+        cache.db so the TUI shot, the dashboard shots, and the report/
+        forecast SVGs all tell the same story (current-week 53% used,
+        $28.62 spent, in-progress Thursday with a WARN ~104% projection).
+
+        Mirrors the 8-week trend table seeded by build-readme-fixtures.py
+        (`_populate_weeks`) so the TUI's Trend panel shows the same
+        $/1% arc the dashboard's Trend modal does.
+
+        Dev-only — production code paths never invoke this. Kept here so
+        the `DataSnapshot` shape stays the single source of truth (mirror
+        any future field additions to keep marketing renders in sync).
+        """
+        as_of = dt.datetime.strptime(
+            as_of_iso, "%Y-%m-%dT%H:%M:%SZ"
+        ).replace(tzinfo=dt.timezone.utc)
+        # Anchor the current subscription week to Monday 00:00 UTC of
+        # `as_of`'s containing week so the marketing copy ("week of …")
+        # lines up with the stats.db rows seeded by build-readme-fixtures.
+        week_start = (as_of - dt.timedelta(days=as_of.weekday())).replace(
+            hour=0, minute=0, second=0, microsecond=0,
+        )
+        week_end = week_start + dt.timedelta(days=7)
+        used_pct = 53.0
+        spent_usd = 28.62
+        cw = TuiCurrentWeek(
+            week_start_at=week_start,
+            week_end_at=week_end,
+            used_pct=used_pct,
+            five_hour_pct=36.0,
+            five_hour_resets_at=as_of.replace(minute=0, second=0, microsecond=0)
+                + dt.timedelta(hours=3),
+            spent_usd=spent_usd,
+            dollars_per_percent=spent_usd / used_pct,
+            latest_snapshot_at=as_of,
+            freshness_label="fresh",
+            freshness_age=12,
+            five_hour_block=None,
+        )
+        # ---- Forecast: WARN, ~98% projected, fits within modal width.
+        # The TUI's verdict mapping (`_tui_verdict_of`) reads
+        # `final_percent_high >= 100` as OVER, `>= 90` as WARN. We want
+        # WARN here, so synthesize r_avg/r_recent that land both
+        # projection bars in the 90s. Note: this is the TUI-only render
+        # path; the dashboard re-derives forecast from the seeded
+        # fixture DB via `snapshot_to_envelope` and lands at ~103%
+        # there (which the dashboard's verdict map calls WARN already
+        # via the `cap` enum, with no >=100 threshold split).
+        elapsed_hours = (as_of - week_start).total_seconds() / 3600.0
+        remaining_hours = max(0.0, (week_end - as_of).total_seconds() / 3600.0)
+        remaining_days = remaining_hours / 24.0
+        # Headline projection target: ~98% → r_avg = (98 - 53) / 82 ≈ 0.549.
+        r_avg = (98.0 - used_pct) / remaining_hours if remaining_hours > 0 else 0.0
+        # Recent 24h slightly lower: ~94% → r_recent = (94 - 53) / 82 ≈ 0.500.
+        r_recent = (94.0 - used_pct) / remaining_hours if remaining_hours > 0 else 0.0
+        p_24h_ago = max(0.0, used_pct - r_recent * 24.0)
+        dpp = spent_usd / used_pct
+        final_low = used_pct + r_recent * remaining_hours
+        final_high = used_pct + r_avg * remaining_hours
+        # Two BudgetRows mirroring the TUI's hard-coded targets [100, 90].
+        budgets = [
+            BudgetRow(
+                target_percent=100,
+                pct_headroom=100.0 - used_pct,
+                dollars_per_day=((100.0 - used_pct) * dpp / remaining_days)
+                                 if remaining_days > 0 else None,
+                percent_per_day=((100.0 - used_pct) / remaining_days)
+                                 if remaining_days > 0 else None,
+            ),
+            BudgetRow(
+                target_percent=90,
+                pct_headroom=90.0 - used_pct,
+                dollars_per_day=((90.0 - used_pct) * dpp / remaining_days)
+                                 if remaining_days > 0 else None,
+                percent_per_day=((90.0 - used_pct) / remaining_days)
+                                 if remaining_days > 0 else None,
+            ),
+        ]
+        forecast_inputs = ForecastInputs(
+            now_utc=as_of,
+            week_start_at=week_start,
+            week_end_at=week_end,
+            elapsed_hours=elapsed_hours,
+            elapsed_fraction=elapsed_hours / 168.0,
+            remaining_hours=remaining_hours,
+            remaining_days=remaining_days,
+            p_now=used_pct,
+            five_hour_percent=36.0,
+            spent_usd=spent_usd,
+            snapshot_count=12,
+            latest_snapshot_at=as_of,
+            p_24h_ago=p_24h_ago,
+            t_24h_actual_hours=24.0,
+            dollars_per_percent=dpp,
+            dollars_per_percent_source="this_week",
+            confidence="high",
+            low_confidence_reasons=[],
+        )
+        forecast = ForecastOutput(
+            inputs=forecast_inputs,
+            r_avg=r_avg,
+            r_recent=r_recent,
+            final_percent_low=final_low,
+            final_percent_high=final_high,
+            projected_cap=final_high >= 100.0,
+            already_capped=False,
+            cap_at=None,
+            budgets=budgets,
+        )
+        # ---- Trend: 8 weeks oldest-first, mirroring the
+        # `_populate_weeks` series in bin/build-readme-fixtures.py.
+        # Spark heights computed the same way `_tui_build_trend` does
+        # (normalize $/1% to 1..8 across the window).
+        weekly_series = [
+            (38.0, 24.70),
+            (41.0, 25.83),
+            (44.0, 25.96),
+            (47.0, 24.91),
+            (50.0, 25.00),
+            (53.0, 22.79),
+            (56.0, 25.20),
+            (used_pct, spent_usd),  # current week
+        ]
+        dpps = [round(c / p, 4) for p, c in weekly_series]
+        lo, hi = min(dpps), max(dpps)
+        span = (hi - lo) or 1e-9
+        trend: list[TuiTrendRow] = []
+        prev_dpp: float | None = None
+        for i, ((pct, cost), wd) in enumerate(zip(weekly_series, dpps)):
+            offset = 7 - i
+            wstart_dt = week_start - dt.timedelta(days=7 * offset)
+            spark = max(1, min(8, int(round((wd - lo) / span * 7)) + 1))
+            delta = (wd - prev_dpp) if prev_dpp is not None else None
+            trend.append(TuiTrendRow(
+                week_label=wstart_dt.strftime("%b %d"),
+                week_start_at=wstart_dt,
+                used_pct=pct,
+                dollars_per_percent=wd,
+                delta_dpp=delta,
+                spark_height=spark,
+                is_current=(i == 7),
+            ))
+            prev_dpp = wd
+        # ---- Sessions: 6 recent rows spanning 4 projects + 3 models,
+        # ordered last-activity desc (matches the aggregator's natural
+        # output, which the TUI's default sort preserves).
+        sessions = [
+            TuiSessionRow(
+                started_at=as_of - dt.timedelta(hours=1, minutes=22),
+                duration_minutes=46.0,
+                model_primary="claude-sonnet-4-6",
+                cost_usd=2.84,
+                cache_hit_pct=87.5,
+                project_label="web-app",
+                session_id="sess-web-app-00",
+            ),
+            TuiSessionRow(
+                started_at=as_of - dt.timedelta(hours=3, minutes=10),
+                duration_minutes=72.0,
+                model_primary="claude-opus-4-7",
+                cost_usd=4.97,
+                cache_hit_pct=72.0,
+                project_label="api-gateway",
+                session_id="sess-api-gateway-01",
+            ),
+            TuiSessionRow(
+                started_at=as_of - dt.timedelta(hours=5, minutes=44),
+                duration_minutes=33.0,
+                model_primary="claude-haiku-4-5-20251001",
+                cost_usd=0.62,
+                cache_hit_pct=91.3,
+                project_label="data-pipeline",
+                session_id="sess-data-pipeline-02",
+            ),
+            TuiSessionRow(
+                started_at=as_of - dt.timedelta(hours=8, minutes=5),
+                duration_minutes=58.0,
+                model_primary="claude-sonnet-4-6",
+                cost_usd=3.41,
+                cache_hit_pct=79.8,
+                project_label="mobile-client",
+                session_id="sess-mobile-client-00",
+            ),
+            TuiSessionRow(
+                started_at=as_of - dt.timedelta(days=1, hours=2),
+                duration_minutes=104.0,
+                model_primary="claude-opus-4-7",
+                cost_usd=6.18,
+                cache_hit_pct=68.4,
+                project_label="web-app",
+                session_id="sess-web-app-01",
+            ),
+            TuiSessionRow(
+                started_at=as_of - dt.timedelta(days=1, hours=6, minutes=30),
+                duration_minutes=29.0,
+                model_primary="claude-sonnet-4-6",
+                cost_usd=1.55,
+                cache_hit_pct=84.1,
+                project_label="api-gateway",
+                session_id="sess-api-gateway-02",
+            ),
+        ]
+        return cls(
+            current_week=cw,
+            forecast=forecast,
+            trend=trend,
+            sessions=sessions,
+            last_sync_at=None,
+            last_sync_error=None,
+            generated_at=as_of,
+        )
+
+
+@dataclass
+class RuntimeState:
+    """Main-thread-only UI state. Not shared with sync thread."""
+    variant: str                  # 'conventional' | 'expressive'
+    focus_index: int              # 0..3 for A; always 3 (sessions) for B
+    session_scroll: int           # topmost visible session row index
+    show_help: bool
+    toast: tuple[str, float] | None   # (message, monotonic_expiry)
+    color_enabled: bool
+    tz: str                        # 'utc' | 'local' | IANA name (legacy token; F4 moved _tui_format_started to consume display_tz directly. Field retained for back-compat call sites.)
+    # Resolved display timezone (per spec §2: --tz flag > config.display.tz > host).
+    # ZoneInfo means "render in this zone"; None means "host-local via bare
+    # astimezone()". Threaded through renderers that call format_display_dt.
+    display_tz: "ZoneInfo | None" = None
+    # ---- v2 additions (spec §3.5, §4.4) ----
+    sort_key: str = "last-activity"      # 'last-activity'|'cost'|'duration'|'model'|'project'
+    filter_term: str | None = None        # None = no active filter
+    search_term: str | None = None        # None = no search; "" = active but empty buffer
+    search_matches: list[int] = field(default_factory=list)  # indices into post-filter+sort list
+    search_index: int = 0                 # current match in search_matches[]
+    input_mode: str | None = None         # 'filter' | 'search' | None
+    input_buffer: str = ""                # live typing during input mode
+    modal_kind: str | None = None         # 'current_week'|'forecast'|'trend'|'session'|None
+    modal_scroll: int = 0                 # topmost visible modal content line
+    # One-shot "snap to bottom on first render" flag for modals that default to
+    # the newest rows (trend, current_week). Set by modal openers; cleared by
+    # the first builder call that performs the snap. Avoids reusing
+    # modal_scroll==0 as a sentinel — otherwise scrolling to the top would
+    # bounce the view back to the bottom on the next redraw.
+    modal_snap_pending: bool = False
+    # ---- v2.4.4 fixture-injection hook (dev-only) ----
+    session_detail_override: Any = None   # TuiSessionDetail | None — injected by fixtures only
+    # Memoized session detail to avoid rebuilding (365-day rescan + re-aggregate)
+    # on every modal redraw tick. Key: (session_id, snap.generated_at).
+    session_detail_cache: Any = None      # tuple[str, dt.datetime, TuiSessionDetail | None] | None
+
+    @classmethod
+    def initial(cls, args) -> "RuntimeState":
+        no_color_env = "NO_COLOR" in os.environ
+        return cls(
+            variant=args.variant,
+            focus_index=3,        # sessions focused by default (design choice)
+            session_scroll=0,
+            show_help=False,
+            toast=None,
+            color_enabled=not (args.no_color or no_color_env),
+            tz=args.tz,
+            display_tz=getattr(args, "_resolved_tz", None),
+        )
+
+
+def _tui_build_current_week(
+    conn: sqlite3.Connection,
+    now_utc: dt.datetime,
+    *,
+    skip_sync: bool = False,
+) -> TuiCurrentWeek | None:
+    """Build the TuiCurrentWeek from the latest snapshot + live cost.
+
+    Returns None when no current-week usage snapshot exists.
+    """
+    fetched = _fetch_current_week_snapshots(conn, now_utc)
+    if fetched is None:
+        return None
+    week_start_at, week_end_at, samples = fetched
+    if not samples:
+        return None
+    # Mirror the reset override applied by `_load_forecast_inputs` so the
+    # Current Week card's spent_usd and $/1% reflect the post-reset window.
+    week_start_at, samples = _apply_midweek_reset_override(
+        conn, week_start_at, week_end_at, samples
+    )
+    if not samples:
+        return None
+    # samples tuple shape: (captured_at_utc, weekly_percent, five_hour_percent).
+    # See _fetch_current_week_snapshots at bin/cctally:9122
+    # (lines ~9189-9194 and ~9221-9226). That helper does not surface
+    # five_hour_resets_at, so do a targeted lookup here for the freshest
+    # non-NULL reset timestamp on the current week.
+    latest = samples[-1]
+    used_pct = float(latest[1])
+    five_hr_pct = float(latest[2]) if latest[2] is not None else None
+    spent = _sum_cost_for_range(
+        week_start_at, now_utc, mode="auto", skip_sync=skip_sync
+    )
+    dpp = (spent / used_pct) if used_pct > 0 else None
+    # Collect every textual variant of week_start_at that parses to the same
+    # instant — mirrors `_fetch_current_week_snapshots` lines 9199-9210 so
+    # legacy local-offset rows and newly UTC-canonicalized rows both contribute.
+    ws_texts = conn.execute(
+        "SELECT DISTINCT week_start_at FROM weekly_usage_snapshots "
+        "WHERE week_start_at IS NOT NULL"
+    ).fetchall()
+    matching_ws_texts: list[str] = []
+    for r in ws_texts:
+        try:
+            rws = parse_iso_datetime(r[0], "week_start_at")
+        except ValueError:
+            continue
+        if rws == week_start_at:
+            matching_ws_texts.append(r[0])
+    five_hr_resets_at: dt.datetime | None = None
+    if matching_ws_texts:
+        placeholders = ",".join("?" * len(matching_ws_texts))
+        reset_row = conn.execute(
+            f"SELECT five_hour_resets_at FROM weekly_usage_snapshots "
+            f"WHERE week_start_at IN ({placeholders}) "
+            f"  AND five_hour_resets_at IS NOT NULL "
+            f"ORDER BY captured_at_utc DESC, id DESC LIMIT 1",
+            tuple(matching_ws_texts),
+        ).fetchone()
+        if reset_row is not None:
+            try:
+                five_hr_resets_at = parse_iso_datetime(
+                    reset_row[0], "five_hour_resets_at"
+                )
+            except ValueError:
+                five_hr_resets_at = None
+            # Suppress stale resets that have already elapsed so renderers
+            # don't show "resets 0h 00m" or a negative duration at the boundary.
+            if five_hr_resets_at is not None and five_hr_resets_at <= now_utc:
+                five_hr_resets_at = None
+    # Freshness — compute label/age from latest snapshot vs. now using the
+    # configured oauth_usage thresholds. Mirrors the dashboard envelope's
+    # cw_freshness derivation in `snapshot_to_envelope`. Refs spec §3.4.
+    captured = latest[0]
+    if isinstance(captured, dt.datetime):
+        if captured.tzinfo is None:
+            captured = captured.replace(tzinfo=dt.timezone.utc)
+        age_s = max(0.0, (now_utc - captured).total_seconds())
+        try:
+            _fresh_cfg = _get_oauth_usage_config(load_config())
+        except Exception:
+            _fresh_cfg = _OAUTH_USAGE_DEFAULTS
+        freshness_label = _freshness_label(age_s, _fresh_cfg)
+        freshness_age = int(age_s)
+    else:
+        freshness_label = None
+        freshness_age = None
+    return TuiCurrentWeek(
+        week_start_at=week_start_at,
+        week_end_at=week_end_at,
+        used_pct=used_pct,
+        five_hour_pct=five_hr_pct,
+        five_hour_resets_at=five_hr_resets_at,
+        spent_usd=float(spent),
+        dollars_per_percent=dpp,
+        latest_snapshot_at=latest[0],
+        freshness_label=freshness_label,
+        freshness_age=freshness_age,
+        five_hour_block=_select_current_block_for_envelope(
+            conn, current_used_pct=used_pct, now_utc=now_utc,
+        ),
+    )
+
+
+def _tui_build_forecast(
+    conn: sqlite3.Connection,
+    now_utc: dt.datetime,
+    *,
+    skip_sync: bool = False,
+):
+    """Call into existing forecast internals. Returns a ForecastOutput or None."""
+    inputs = _load_forecast_inputs(conn, now_utc, skip_sync=skip_sync)
+    if inputs is None:
+        return None
+    return _compute_forecast(inputs, [100, 90])
+
+
+def _tui_build_trend(
+    conn: sqlite3.Connection,
+    now_utc: dt.datetime,
+    *,
+    skip_sync: bool = False,  # noqa: ARG001 — unused today, kept for API symmetry
+    count: int = 8,
+    display_tz: "ZoneInfo | None" = None,
+) -> list[TuiTrendRow]:
+    """Build the last `count` trend rows, chronological (oldest first).
+
+    `cmd_report` inlines its row build rather than delegating to a helper,
+    so instead of refactoring the subcommand we call the same underlying
+    loaders (`get_recent_weeks` + `get_latest_usage_for_week` +
+    `get_latest_cost_for_week`) directly here. Output for the shared
+    columns (`week_start_at`, `used_pct`, `dollars_per_percent`) matches
+    `cmd_report` byte-for-byte — verified in the bundle regression diff.
+    """
+    # `get_recent_weeks` returns WeekRef rows DESC by week_start_date,
+    # already routed through `_apply_reset_events_to_weekrefs` so credited
+    # weeks come back as TWO refs (pre-credit + post-credit) sharing the
+    # same `WeekRef.key`.
+    week_refs = get_recent_weeks(conn, max(1, count))
+
+    # Figure out which week_ref corresponds to the current subscription week.
+    # Mirrors `cmd_report`'s Bug D pattern: build a current_ref from the
+    # latest usage snapshot, route it through `_apply_reset_events_to_weekrefs`
+    # so its `week_start_at` reflects the post-credit segment (or the
+    # original start for non-credit weeks), then disambiguate the
+    # synthesized pre-credit ref from the live post-credit ref via BOTH
+    # `key` AND `week_start_at`. Key-only equality marks both segments
+    # as current, which is why the dashboard's trend panel previously
+    # showed two adjacent rows both highlighted as "current" with
+    # identical 4.0% values on the user's live in-place credit data.
+    latest_usage = conn.execute(
+        "SELECT week_start_date, week_end_date "
+        "FROM weekly_usage_snapshots "
+        "ORDER BY captured_at_utc DESC, id DESC LIMIT 1"
+    ).fetchone()
+    current_key: str | None = None
+    current_week_start_at: str | None = None
+    if latest_usage is not None and latest_usage["week_start_date"] is not None:
+        current_key = latest_usage["week_start_date"]
+        try:
+            c = _cctally()
+            canon_start, canon_end = c._get_canonical_boundary_for_date(
+                conn, latest_usage["week_start_date"]
+            )
+            current_ref = c.make_week_ref(
+                week_start_date=latest_usage["week_start_date"],
+                week_end_date=latest_usage["week_end_date"],
+                week_start_at=canon_start,
+                week_end_at=canon_end,
+            )
+            _adjusted = c._apply_reset_events_to_weekrefs(conn, [current_ref])
+            if _adjusted:
+                current_week_start_at = _adjusted[0].week_start_at
+        except (ValueError, sqlite3.DatabaseError, AttributeError):
+            current_week_start_at = None
+
+    # Build an intermediate list of (week_ref, used_pct, dpp) in oldest-first
+    # chronological order.
+    chrono = list(reversed(week_refs))
+    # Split-key set (Bug D): credited weeks appear twice in `week_refs`
+    # with identical `WeekRef.key`. For those keys ONLY, pin
+    # `as_of_utc=week_ref.week_end_at` so each segment finds its own
+    # latest snapshot — without this both segments collapse to the
+    # post-credit snapshot's weekly_percent. Non-credit weeks (single
+    # ref per key) keep the legacy unfiltered lookup.
+    _split_keys = {
+        r.key
+        for r in week_refs
+        if sum(1 for x in week_refs if x.key == r.key) > 1
+    }
+    intermediate: list[tuple[Any, float | None, float | None]] = []
+    for week_ref in chrono:
+        usage = get_latest_usage_for_week(
+            conn,
+            week_ref,
+            as_of_utc=(
+                week_ref.week_end_at if week_ref.key in _split_keys else None
+            ),
+        )
+        # See cmd_report for why reset-affected weeks skip the cost cache
+        # and live-compute from session_entries over the effective range.
+        if _week_ref_has_reset_event(conn, week_ref):
+            cost_usd = _compute_cost_for_weekref(week_ref)
+        else:
+            cost = get_latest_cost_for_week(conn, week_ref)
+            cost_usd = float(cost["cost_usd"]) if cost else None
+        percent = float(usage["weekly_percent"]) if usage else None
+        ratio = (cost_usd / percent) if (
+            cost_usd is not None and percent and percent > 0
+        ) else None
+        intermediate.append((week_ref, percent, ratio))
+
+    # Normalize dpp into spark heights 1..8 across the window.
+    dpps = [d for _, _, d in intermediate if d is not None]
+    if dpps:
+        lo, hi = min(dpps), max(dpps)
+        span = (hi - lo) or 1e-9
+    else:
+        lo, hi, span = 0.0, 1.0, 1e-9
+
+    out: list[TuiTrendRow] = []
+    prev_dpp: float | None = None
+    for week_ref, percent, dpp in intermediate:
+        delta = (dpp - prev_dpp) if (dpp is not None and prev_dpp is not None) else None
+        spark = 1
+        if dpp is not None:
+            spark = int(round((dpp - lo) / span * 7)) + 1
+            spark = max(1, min(8, spark))
+        # WeekRef.week_start is a date; synthesize a UTC datetime so
+        # TuiTrendRow carries a timezone-aware instant (prefer the explicit
+        # week_start_at if present).
+        if week_ref.week_start_at:
+            week_start_dt = parse_iso_datetime(
+                week_ref.week_start_at, "week_start_at"
+            )
+            week_label = format_display_dt(
+                week_start_dt, display_tz, fmt="%b %d", suffix=False,
+            )
+        else:
+            week_start_dt = dt.datetime.combine(
+                week_ref.week_start, dt.time(0, 0), dt.timezone.utc
+            )
+            # No real boundary instant — format the calendar date directly so
+            # localizing midnight-UTC doesn't shift it to the prior day in
+            # zones west of UTC (e.g. 2026-04-14 → "Apr 13" in America/New_York).
+            week_label = week_ref.week_start.strftime("%b %d")
+        # Bug G (v1.7.2 round-5): match on BOTH `key` AND `week_start_at`
+        # for credited weeks so the pre-credit synthesized ref doesn't
+        # also light up as "current" — both refs share `key`, only their
+        # `week_start_at` differs (post-credit = effective reset moment,
+        # pre-credit = original API-derived start). Non-credit weeks
+        # have only one ref per key so `week_start_at` matching is
+        # automatic. When `current_week_start_at` is None (no reset
+        # event for the current week, or the resolution above failed),
+        # falls back to legacy key-only matching.
+        is_cur = (
+            current_key is not None
+            and week_ref.key == current_key
+            and (
+                current_week_start_at is None
+                or week_ref.week_start_at == current_week_start_at
+            )
+        )
+        out.append(TuiTrendRow(
+            week_label=week_label,
+            week_start_at=week_start_dt,
+            # Preserve None when no usage snapshot exists for this week —
+            # matches `cmd_report`'s "n/a" rendering (9980) and avoids
+            # fabricating a 0.0% row for the phantom-week case (cost
+            # snapshot present, usage snapshot absent).
+            used_pct=float(percent) if percent is not None else None,
+            dollars_per_percent=dpp,
+            delta_dpp=delta,
+            spark_height=spark,
+            is_current=is_cur,
+        ))
+        if dpp is not None:
+            prev_dpp = dpp
+    return out
+
+
+def _tui_build_weekly_history(
+    conn: sqlite3.Connection,
+    now_utc: dt.datetime,
+    *,
+    skip_sync: bool = False,
+    count: int = 12,
+    display_tz: "ZoneInfo | None" = None,
+) -> list[TuiTrendRow]:
+    """Return the last `count` weeks for the Trend modal (spec §4.6.3).
+
+    Same data shape as `_tui_build_trend` (the panel) — just more rows.
+    The panel renders 8; the modal renders up to 12. Wrapping rather
+    than parameterising the call site keeps the snapshot fields
+    semantically distinct (panel data vs. modal data) and avoids
+    accidental cross-contamination.
+    """
+    return _tui_build_trend(
+        conn, now_utc, skip_sync=skip_sync, count=count, display_tz=display_tz,
+    )
+
+
+def _tui_build_sessions(
+    now_utc: dt.datetime,
+    *,
+    limit: int = 100,
+    skip_sync: bool = False,
+) -> list[TuiSessionRow]:
+    """Load the last `limit` Claude sessions (merged across resumes).
+
+    Started-time descending (matches `_aggregate_claude_sessions` —
+    sorted by `last_activity` DESC). Uses the same aggregator as the
+    `session` subcommand, so row identity and project labels match
+    `cctally session --json` exactly.
+
+    When `skip_sync=True`, honors the parent's `--no-sync` intent: no
+    ingest pass, just read whatever is already cached.
+    """
+    # Bounded scan window — the sessions pane promises "last `limit`". A
+    # 365-day scan covers virtually all users (even one-session-every-few-days
+    # sparseness still nets the cap). Bounded rather than all-history so
+    # sync-tick cost stays predictable on heavy DBs: the aggregator runs
+    # on every entry in the window before slicing.
+    range_start = now_utc - dt.timedelta(days=365)
+    entries = get_claude_session_entries(range_start, now_utc, skip_sync=skip_sync)
+    sessions = _aggregate_claude_sessions(entries)   # last_activity desc
+    out: list[TuiSessionRow] = []
+    for s in sessions[:limit]:
+        duration_min = (s.last_activity - s.first_activity).total_seconds() / 60.0
+        total_read = s.cache_read_tokens
+        total_io = s.input_tokens + s.cache_creation_tokens + s.cache_read_tokens
+        cache_pct = (total_read / total_io * 100) if total_io > 0 else None
+        out.append(TuiSessionRow(
+            started_at=s.first_activity,
+            duration_minutes=duration_min,
+            model_primary=(s.models[0] if s.models else "—"),
+            cost_usd=s.cost_usd,
+            cache_hit_pct=cache_pct,
+            project_label=os.path.basename(s.project_path) or s.project_path,
+            session_id=s.session_id,
+        ))
+    return out
+
+
+@dataclass
+class TuiSessionDetail:
+    """Detailed view for the Session modal (spec §4.6.4).
+
+    Built on demand when the modal opens — not part of DataSnapshot.
+    """
+    session_id: str
+    started_at: dt.datetime
+    last_activity_at: dt.datetime
+    duration_minutes: float
+    project_label: str
+    project_path: str                     # full cwd
+    source_paths: list[str]               # JSONL file paths (for resumed sessions, may be >1)
+    models: list[tuple[str, str]]         # [(model_name, role)] role in {"primary","secondary"}
+    input_tokens: int
+    cache_creation_tokens: int
+    cache_read_tokens: int
+    output_tokens: int
+    cache_hit_pct: float | None
+    cost_per_model: list[tuple[str, float]]   # [(model_name, cost_usd)]
+    cost_total_usd: float
+
+
+def _tui_build_session_detail(
+    session_id: str,
+    *,
+    now_utc: dt.datetime | None = None,
+) -> TuiSessionDetail | None:
+    """Look up one session by ID; return None if not found.
+
+    Reuses the same `get_claude_session_entries` + `_aggregate_claude_sessions`
+    pipeline as `_tui_build_sessions` but filters down to the matching ID.
+    Bounded scan window matches the panel builder (365 days).
+    """
+    now_utc = now_utc or dt.datetime.now(dt.timezone.utc)
+    range_start = now_utc - dt.timedelta(days=365)
+    entries = get_claude_session_entries(range_start, now_utc, skip_sync=True)
+    sessions = _aggregate_claude_sessions(entries)
+    match: Any | None = None
+    for s in sessions:
+        if s.session_id == session_id:
+            match = s
+            break
+    if match is None:
+        return None
+    duration_min = (match.last_activity - match.first_activity).total_seconds() / 60.0
+    total_read = match.cache_read_tokens
+    total_io = match.input_tokens + match.cache_creation_tokens + match.cache_read_tokens
+    cache_pct = (total_read / total_io * 100) if total_io > 0 else None
+    # Build per-model rows. Role is "primary" for the first model seen in
+    # this session (matches `_aggregate_claude_sessions` `models_order`),
+    # "secondary" for the rest.
+    models_with_role: list[tuple[str, str]] = []
+    for i, m in enumerate(match.models):
+        models_with_role.append((m, "primary" if i == 0 else "secondary"))
+    # Per-model cost: prefer the aggregator's `model_breakdowns` (list of
+    # dicts with `"model"` / `"cost"` keys, sorted by cost desc). Fall back
+    # defensively to a single-row total if the attribute is missing or empty
+    # so the modal stays renderable on any aggregator shape change.
+    cost_per_model: list[tuple[str, float]] = []
+    breakdowns = getattr(match, "model_breakdowns", None)
+    if isinstance(breakdowns, list) and breakdowns:
+        for mb in breakdowns:
+            try:
+                cost_per_model.append((str(mb["model"]), float(mb["cost"])))
+            except (KeyError, TypeError, ValueError):
+                continue
+    if not cost_per_model and match.models:
+        # Single-model fallback: attribute total to primary.
+        cost_per_model.append((match.models[0], float(match.cost_usd)))
+    return TuiSessionDetail(
+        session_id=match.session_id,
+        started_at=match.first_activity,
+        last_activity_at=match.last_activity,
+        duration_minutes=duration_min,
+        project_label=os.path.basename(match.project_path) or match.project_path,
+        project_path=match.project_path,
+        source_paths=list(match.source_paths or []),
+        models=models_with_role,
+        input_tokens=int(match.input_tokens),
+        cache_creation_tokens=int(match.cache_creation_tokens),
+        cache_read_tokens=int(match.cache_read_tokens),
+        output_tokens=int(match.output_tokens),
+        cache_hit_pct=cache_pct,
+        cost_per_model=cost_per_model,
+        cost_total_usd=float(match.cost_usd),
+    )
+
+
+def _tui_build_snapshot(
+    *,
+    now_utc: dt.datetime | None = None,
+    skip_sync: bool = False,
+    display_tz_pref_override: "str | None" = None,
+) -> DataSnapshot:
+    """Single-shot build of a DataSnapshot from the DB + cache.
+
+    Runs in the sync thread. Catches exceptions per sub-build and records
+    them on `last_sync_error` so the UI can surface them without crashing.
+
+    ``display_tz_pref_override`` (F3): a canonical tz token (``"local"``
+    / ``"utc"`` / IANA name) that overrides ``config.display.tz`` for
+    the lifetime of this build. Used by ``cmd_dashboard`` when ``--tz``
+    is supplied so the in-memory zone wins over the persisted config
+    without modifying it. ``None`` means "respect config".
+    """
+    import time
+    now_utc = now_utc or dt.datetime.now(dt.timezone.utc)
+    # Resolve the display tz once per snapshot so labels rendered into
+    # BlocksPanelRow / future panel rows share a single zone with the
+    # envelope's `display` block. Routed through the shared
+    # `_resolve_display_tz_obj` helper so this site, `_compute_display_block`,
+    # and `_handle_get_block_detail` share identical fallback semantics
+    # (one-shot stderr warning on local-resolution failure). Not threaded
+    # into label-FOR-LOOKUP paths like `_aggregate_monthly` keys (out of
+    # scope for Task 11).
+    _build_display_tz = _resolve_display_tz_obj(
+        _apply_display_tz_override(load_config(), display_tz_pref_override)
+    )
+    conn = open_db()
+    try:
+        errors: list[str] = []
+        cw: TuiCurrentWeek | None = None
+        fc: Any | None = None
+        trend: list[TuiTrendRow] = []
+        sessions: list[TuiSessionRow] = []
+        milestones: list[TuiPercentMilestone] = []
+        history: list[TuiTrendRow] = []
+        weekly_periods: list[WeeklyPeriodRow] = []
+        monthly_periods: list[MonthlyPeriodRow] = []
+        blocks_panel: list[BlocksPanelRow] = []
+        daily_panel:  list[DailyPanelRow]  = []
+        alerts: list[dict] = []
+        try:
+            cw = _tui_build_current_week(conn, now_utc, skip_sync=skip_sync)
+        except Exception as exc:
+            errors.append(f"current-week: {exc}")
+        try:
+            fc = _tui_build_forecast(conn, now_utc, skip_sync=skip_sync)
+        except Exception as exc:
+            errors.append(f"forecast: {exc}")
+        try:
+            trend = _tui_build_trend(
+                conn, now_utc, skip_sync=skip_sync, display_tz=_build_display_tz,
+            )
+        except Exception as exc:
+            errors.append(f"trend: {exc}")
+        try:
+            # The sessions aggregator goes through
+            # `get_claude_session_entries`, which runs `sync_cache` unless
+            # `skip_sync=True` is threaded through. Honor the caller's
+            # intent so `--no-sync` and the initial cache-only paint
+            # both avoid ingest latency/lock contention.
+            sessions = _tui_build_sessions(now_utc, skip_sync=skip_sync)
+        except Exception as exc:
+            errors.append(f"sessions: {exc}")
+        # ---- v2 additions ----
+        try:
+            if cw is not None:
+                milestones = _tui_build_percent_milestones(conn)
+        except Exception as exc:
+            errors.append(f"milestones: {exc}")
+        try:
+            history = _tui_build_weekly_history(
+                conn, now_utc, skip_sync=skip_sync, display_tz=_build_display_tz,
+            )
+        except Exception as exc:
+            errors.append(f"weekly-history: {exc}")
+        # ---- v2.1 additions: dashboard Weekly / Monthly panels ----
+        try:
+            weekly_periods = _dashboard_build_weekly_periods(
+                conn, now_utc, n=12, skip_sync=skip_sync
+            )
+        except Exception as exc:
+            errors.append(f"weekly-periods: {exc}")
+        try:
+            monthly_periods = _dashboard_build_monthly_periods(
+                conn, now_utc, n=12, skip_sync=skip_sync,
+                display_tz=_build_display_tz,
+            )
+        except Exception as exc:
+            errors.append(f"monthly-periods: {exc}")
+        # ---- v2.2 additions: dashboard Blocks / Daily panels ----
+        try:
+            if cw is not None:
+                blocks_panel = _dashboard_build_blocks_panel(
+                    conn, now_utc,
+                    week_start_at=cw.week_start_at,
+                    week_end_at=cw.week_end_at,
+                    skip_sync=skip_sync,
+                    display_tz=_build_display_tz,
+                )
+        except Exception as exc:
+            errors.append(f"blocks-panel: {exc}")
+        try:
+            daily_panel = _dashboard_build_daily_panel(
+                conn, now_utc, n=30, skip_sync=skip_sync,
+                display_tz=_build_display_tz,
+            )
+        except Exception as exc:
+            errors.append(f"daily-panel: {exc}")
+        # ---- threshold-actions T5: alerts envelope array ----
+        # Precomputed at sync time so `snapshot_to_envelope` stays a pure
+        # renderer (no DB I/O on the dashboard hot path; mirrors how
+        # `current_week.five_hour_block` is precomputed via
+        # `_select_current_block_for_envelope`).
+        try:
+            alerts = _build_alerts_envelope_array(conn)
+        except Exception as exc:
+            errors.append(f"alerts: {exc}")
+        return DataSnapshot(
+            current_week=cw,
+            forecast=fc,
+            trend=trend,
+            sessions=sessions,
+            last_sync_at=time.monotonic(),
+            last_sync_error=("; ".join(errors) if errors else None),
+            generated_at=now_utc,
+            percent_milestones=milestones,
+            weekly_history=history,
+            weekly_periods=weekly_periods,
+            monthly_periods=monthly_periods,
+            blocks_panel=blocks_panel,
+            daily_panel=daily_panel,
+            alerts=alerts,
+        )
+    finally:
+        conn.close()
+
+
+def _tui_empty_snapshot(now_utc: dt.datetime) -> DataSnapshot:
+    """First-paint placeholder with no data loaded yet."""
+    return DataSnapshot(
+        current_week=None, forecast=None, trend=[], sessions=[],
+        last_sync_at=None, last_sync_error=None, generated_at=now_utc,
+        percent_milestones=[], weekly_history=[],
+        weekly_periods=[], monthly_periods=[],
+        blocks_panel=[], daily_panel=[],
+    )
+
+
+class TuiKeyReader:
+    """Context manager for raw-mode stdin reads.
+
+    Non-TTY input degrades gracefully — read() always returns None.
+    """
+
+    _ESC_MAP = {
+        "[A": "up",    "[B": "down",
+        "[C": "right", "[D": "left",
+        "[5~": "pgup", "[6~": "pgdn",
+        "[H":  "home", "[F":  "end",
+    }
+
+    def __init__(self) -> None:
+        self._fd = None
+        self._saved = None
+
+    def __enter__(self):
+        try:
+            import termios, tty
+        except ImportError:
+            return self  # non-posix: degrade to null reader
+        if not sys.stdin.isatty():
+            return self
+        try:
+            self._fd = sys.stdin.fileno()
+            self._saved = termios.tcgetattr(self._fd)
+            tty.setcbreak(self._fd)
+        except Exception:
+            # Degrade gracefully on any setup error.
+            self._fd = None
+            self._saved = None
+        return self
+
+    def __exit__(self, *exc):
+        if self._fd is not None and self._saved is not None:
+            try:
+                import termios
+                termios.tcsetattr(self._fd, termios.TCSADRAIN, self._saved)
+            except Exception:
+                pass
+
+    # SS3 arrows (ESC O X) — sent by terminals in DECCKM "application" mode.
+    _SS3_MAP = {
+        b"A": "up",    b"B": "down",
+        b"C": "right", b"D": "left",
+        b"H": "home",  b"F": "end",
+    }
+
+    def read(self, timeout: float) -> str | None:
+        """Blocking read up to `timeout` seconds. Returns a key name or char.
+
+        Reads via os.read on the raw fd rather than sys.stdin.read, because
+        the TextIOWrapper on sys.stdin buffers ahead: sys.stdin.read(1) pulls
+        an ESC sequence like b"\\x1b[B" into Python's buffer as a block, and
+        the follow-up select() on the fd then sees nothing and times out —
+        causing the reader to mis-return "esc" and the handler to quit.
+        """
+        import select
+        if not sys.stdin.isatty() or self._fd is None:
+            return None
+        fd = self._fd
+        try:
+            r, _, _ = select.select([fd], [], [], max(0.0, timeout))
+            if not r:
+                return None
+            b = os.read(fd, 1)
+        except Exception:
+            return None
+        if not b:
+            return None
+        if b == b"\x1b":  # ESC; possibly start of CSI or SS3 sequence
+            # 50ms grace to distinguish lone ESC from an arrow sequence.
+            try:
+                r2, _, _ = select.select([fd], [], [], 0.05)
+                if not r2:
+                    return "esc"
+                rest = os.read(fd, 1)
+            except Exception:
+                return "esc"
+            if rest == b"O":
+                # SS3 (application keypad): ESC O X
+                try:
+                    r3, _, _ = select.select([fd], [], [], 0.05)
+                    if not r3:
+                        return "esc"
+                    ch = os.read(fd, 1)
+                except Exception:
+                    return "esc"
+                return self._SS3_MAP.get(ch, None)
+            if rest != b"[":
+                return "esc"
+            seq = b"["
+            # Read until terminator (letter, ~, or 4-char cap).
+            for _ in range(4):
+                try:
+                    r3, _, _ = select.select([fd], [], [], 0.05)
+                    if not r3:
+                        break
+                    ch = os.read(fd, 1)
+                except Exception:
+                    break
+                if not ch:
+                    break
+                seq += ch
+                if ch == b"~" or (b"A" <= ch <= b"Z") or (b"a" <= ch <= b"z"):
+                    break
+            return self._ESC_MAP.get(seq.decode("ascii", errors="replace"), None)
+        if b == b"\t":
+            return "tab"
+        if b == b"\n" or b == b"\r":
+            return "enter"
+        if b == b"\x7f" or b == b"\x08":
+            return "backspace"
+        if b == b"\x03":
+            return "ctrl-c"
+        try:
+            return b.decode("utf-8", errors="replace")
+        except Exception:
+            return None
+
+
+def _tui_handle_key(
+    key: str,
+    runtime: RuntimeState,
+    snapshot_ref: "_SnapshotRef",
+) -> tuple[bool, bool]:
+    """Mutate `runtime` in place. Returns (should_redraw, should_quit).
+
+    `snapshot_ref` is the shared-state holder; key handler may request a
+    force sync via snapshot_ref.request_sync().
+    """
+    import time
+    # Dismiss toast on any key.
+    if runtime.toast is not None:
+        runtime.toast = None
+
+    # v2: modal state — captures most keys (spec §2.3 Modal column).
+    # Placed first so Esc dismisses the modal instead of falling
+    # through to the dashboard's default Esc-quits. modal_kind and
+    # input_mode are mutually exclusive (modal openers gate on
+    # input_mode is None; input openers gate on modal_kind is None),
+    # so this branch never collides with input-mode dispatch.
+    if runtime.modal_kind is not None:
+        if key == "esc":
+            runtime.modal_kind = None
+            runtime.modal_scroll = 0
+            return True, False
+        if key in ("q", "ctrl-c"):
+            return False, True   # quit always works
+        if key in ("up", "k"):
+            runtime.modal_scroll = max(0, runtime.modal_scroll - 1)
+            return True, False
+        if key in ("down", "j"):
+            runtime.modal_scroll = runtime.modal_scroll + 1
+            return True, False
+        if key == "pgup":
+            runtime.modal_scroll = max(0, runtime.modal_scroll - 10)
+            return True, False
+        if key == "pgdn":
+            runtime.modal_scroll = runtime.modal_scroll + 10
+            return True, False
+        # All other dashboard-layer keys (Tab, s, f, /, v, r, ?, Enter, 1-4)
+        # are silently swallowed per spec §2.4.
+        return True, False
+
+    # In input mode, only ctrl-c quits; esc/q are handled by the input
+    # mode dispatch (esc cancels, q is just a printable character to append).
+    if runtime.input_mode is not None and key == "ctrl-c":
+        return False, True
+    if runtime.input_mode is None and key in ("q", "ctrl-c", "esc"):
+        # Esc only quits when no help overlay is showing.
+        if key == "esc" and runtime.show_help:
+            runtime.show_help = False
+            return True, False
+        return False, True
+    if runtime.input_mode is None and key == "?":
+        runtime.show_help = not runtime.show_help
+        return True, False
+    if runtime.input_mode is None and key == "v":
+        runtime.variant = ("expressive" if runtime.variant == "conventional"
+                           else "conventional")
+        return True, False
+    if runtime.input_mode is None and key == "r":
+        snapshot_ref.request_sync()
+        runtime.toast = ("syncing…", time.monotonic() + 1.0)
+        return True, False
+    if runtime.input_mode is None and key == "tab":
+        runtime.focus_index = (runtime.focus_index + 1) % 4
+        return True, False
+    # Scroll (targets sessions when focused; in variant B, always sessions).
+    is_sessions_focus = (runtime.variant == "expressive"
+                         or runtime.focus_index == 3)
+    if runtime.input_mode is None and is_sessions_focus:
+        # v2: n / N — navigate confirmed search matches (spec §3.3).
+        if (key in ("n", "N")
+                and runtime.search_term is not None
+                and runtime.search_matches
+                and runtime.modal_kind is None):
+            if key == "n":
+                runtime.search_index = (runtime.search_index + 1) % len(runtime.search_matches)
+            else:
+                runtime.search_index = (runtime.search_index - 1) % len(runtime.search_matches)
+            runtime.session_scroll = runtime.search_matches[runtime.search_index]
+            return True, False
+        if key in ("up", "k"):
+            runtime.session_scroll = max(0, runtime.session_scroll - 1)
+            return True, False
+        if key in ("down", "j"):
+            runtime.session_scroll = runtime.session_scroll + 1
+            return True, False
+        if key == "pgup":
+            runtime.session_scroll = max(0, runtime.session_scroll - 10)
+            return True, False
+        if key == "pgdn":
+            runtime.session_scroll = runtime.session_scroll + 10
+            return True, False
+    # v2: sessions sort cycle (spec §3.1). Sessions-scoped regardless of focus.
+    if key == "s" and runtime.input_mode is None and runtime.modal_kind is None:
+        runtime.sort_key = _tui_next_sort_key(runtime.sort_key)
+        # Spec §3.3: search clears when sort changes — match indices were
+        # computed against the previous ordering and would jump to wrong rows.
+        runtime.search_term = None
+        runtime.search_matches = []
+        runtime.search_index = 0
+        return True, False
+
+    # v2: filter — open input mode (spec §3.2). Sessions-scoped regardless of focus.
+    if key == "f" and runtime.input_mode is None and runtime.modal_kind is None:
+        runtime.input_mode = "filter"
+        # Edit-existing semantics: pre-load the buffer with current filter.
+        runtime.input_buffer = runtime.filter_term or ""
+        runtime.show_help = False  # mirror Enter/1-4: state change closes help
+        return True, False
+
+    # v2: filter input mode key dispatch (spec §2.3 + §3.2).
+    if runtime.input_mode == "filter":
+        if key == "esc":
+            runtime.input_mode = None
+            runtime.input_buffer = ""
+            return True, False
+        if key == "enter":
+            buf = runtime.input_buffer.strip()
+            runtime.filter_term = buf if buf else None
+            runtime.input_mode = None
+            runtime.input_buffer = ""
+            # Reset session_scroll to top so user lands on first match.
+            runtime.session_scroll = 0
+            # Spec §3.3: search clears when filter changes — narrowing
+            # invalidates the match index list.
+            runtime.search_term = None
+            runtime.search_matches = []
+            runtime.search_index = 0
+            return True, False
+        if key == "backspace":
+            runtime.input_buffer = runtime.input_buffer[:-1]
+            return True, False
+        # Printable: append. Multi-layer defense per memory:
+        # clip on append (max 200), only printable, ignore unrecognised.
+        if isinstance(key, str) and len(key) == 1 and key.isprintable():
+            if len(runtime.input_buffer) < 200:
+                runtime.input_buffer += key
+            return True, False
+        # All other keys swallowed silently in input mode.
+        return True, False
+
+    # v2: search — open input mode (spec §3.3).
+    if key == "/" and runtime.input_mode is None and runtime.modal_kind is None:
+        runtime.input_mode = "search"
+        runtime.input_buffer = ""  # always start fresh per spec §3.3
+        # Stale search_index from a prior query would make the first
+        # post-confirm n/N wrap past the first matches of the new query.
+        runtime.search_index = 0
+        runtime.show_help = False  # mirror Enter/1-4: state change closes help
+        return True, False
+
+    # v2: search input mode key dispatch.
+    if runtime.input_mode == "search":
+        if key == "esc":
+            runtime.input_mode = None
+            runtime.input_buffer = ""
+            # Cancel restores selection: clear matches/highlights.
+            runtime.search_term = None
+            runtime.search_matches = []
+            runtime.search_index = 0
+            return True, False
+        if key == "enter":
+            buf = runtime.input_buffer
+            runtime.search_term = buf if buf else None
+            runtime.input_mode = None
+            runtime.input_buffer = ""
+            # Match list will be populated by the renderer (it knows the
+            # current post-filter+sort list). For now, scroll stays where
+            # the live jump put it.
+            return True, False
+        if key == "backspace":
+            runtime.input_buffer = runtime.input_buffer[:-1]
+            return True, False
+        if isinstance(key, str) and len(key) == 1 and key.isprintable():
+            if len(runtime.input_buffer) < 200:
+                runtime.input_buffer += key
+            return True, False
+        return True, False
+
+    # v2: Enter — open detail modal (spec §2.3 + §4.2).
+    # Both variants: focus_index maps to modal kind.
+    if key == "enter" and runtime.input_mode is None and runtime.modal_kind is None:
+        target_kind = ("current_week", "forecast", "trend", "session")[runtime.focus_index]
+        runtime.modal_kind = target_kind
+        runtime.modal_scroll = 0
+        runtime.modal_snap_pending = True  # trend/current_week: snap to bottom on first render
+        runtime.show_help = False  # mutually exclusive (spec §4.2)
+        return True, False
+
+    # v2: 1-4 universal modal shortcuts (spec §1, Q6a).
+    if (key in ("1", "2", "3", "4")
+            and runtime.input_mode is None
+            and runtime.modal_kind is None):
+        target_kind = ("current_week", "forecast", "trend", "session")[int(key) - 1]
+        runtime.modal_kind = target_kind
+        runtime.modal_scroll = 0
+        runtime.modal_snap_pending = True
+        runtime.show_help = False
+        return True, False
+
+    return False, False
+
+
+class _TuiSyncThread:
+    """Daemon thread that periodically rebuilds the DataSnapshot.
+
+    Honors --no-sync by never syncing (only reading the current DB state).
+    Force-refresh via `snapshot_ref.request_sync()` — thread interrupts its
+    sleep and rebuilds immediately.
+
+    When ``now_utc`` is provided (propagated from cmd_tui when --as-of is
+    set), every rebuild pins the snapshot clock to that value so live mode
+    mirrors --render-once determinism. When None, rebuilds use wall clock.
+    """
+
+    def __init__(
+        self,
+        snapshot_ref: _SnapshotRef,
+        interval: float,
+        *,
+        skip_sync: bool,
+        now_utc: dt.datetime | None = None,
+        display_tz_pref_override: "str | None" = None,
+    ) -> None:
+        import threading
+        self._ref = snapshot_ref
+        self._interval = interval
+        self._skip_sync = skip_sync
+        self._now_utc = now_utc
+        self._display_tz_pref_override = display_tz_pref_override
+        self._stop = threading.Event()
+        self._thread = threading.Thread(target=self._run, daemon=True,
+                                        name="tui-sync")
+
+    def start(self) -> None:
+        self._thread.start()
+
+    def stop(self) -> None:
+        """Signal the thread to exit and wait up to `interval + 0.5s`
+        for it to finish. Because the thread is daemon=True, a failed
+        join will not block process exit."""
+        self._stop.set()
+        self._thread.join(timeout=self._interval + 0.5)
+
+    def _run(self) -> None:
+        import time
+        while not self._stop.is_set():
+            try:
+                # Route through cctally so test monkeypatches on
+                # ``_tui_build_snapshot`` propagate into the sync thread
+                # (cf. _make_run_sync_now_locked above).
+                snap = sys.modules["cctally"]._tui_build_snapshot(
+                    now_utc=self._now_utc, skip_sync=self._skip_sync,
+                    display_tz_pref_override=self._display_tz_pref_override,
+                )
+                self._ref.set(snap)
+            except Exception as exc:
+                # Don't crash the thread on unexpected errors — surface in UI.
+                prev = self._ref.get()
+                self._ref.set(DataSnapshot(
+                    current_week=prev.current_week,
+                    forecast=prev.forecast,
+                    trend=prev.trend,
+                    sessions=prev.sessions,
+                    last_sync_at=prev.last_sync_at,
+                    last_sync_error=f"sync crashed: {exc}",
+                    generated_at=dt.datetime.now(dt.timezone.utc),
+                    percent_milestones=prev.percent_milestones,
+                    weekly_history=prev.weekly_history,
+                    weekly_periods=prev.weekly_periods,
+                    monthly_periods=prev.monthly_periods,
+                    blocks_panel=prev.blocks_panel,
+                    daily_panel=prev.daily_panel,
+                ))
+            # Wait up to interval, or until forced.
+            for _ in range(int(max(1, self._interval * 10))):
+                if self._stop.is_set():
+                    return
+                if self._ref.take_sync_request():
+                    break
+                time.sleep(0.1)
+
+
+def _tui_panel_current_week(
+    snap: DataSnapshot,
+    runtime: RuntimeState,
+    width: int,
+    *,
+    focused: bool,
+) -> list[str]:
+    """Return the list of pre-box body lines (color-tagged) for Variant A.
+
+    Caller wraps in _tui_box_lines and color-tags + recolors the border.
+    Width math is approximate here; the assembler strips tags before
+    measuring.
+    """
+    cw = snap.current_week
+    if cw is None:
+        return [
+            "",
+            "  {dim}no current-week data yet{/}",
+            "  {dim}run `record-usage` to start capturing{/}",
+            "",
+        ]
+    # The panel interior is width - 2. The design uses leftW-20 as bar width.
+    bar_w = max(10, width - 20)
+    used_cls = _tui_bar_color(cw.used_pct)
+    bar_fill = _tui_bar_string(cw.used_pct, bar_w)
+
+    five = cw.five_hour_pct or 0.0
+    five_bar = _tui_bar_string(five, bar_w)
+
+    reset_delta = cw.week_end_at - snap.generated_at
+    reset_days = max(0, reset_delta.days)
+    reset_hrs = max(0, reset_delta.seconds // 3600)
+    snap_age = int((snap.generated_at - cw.latest_snapshot_at).total_seconds())
+    snap_age_m, snap_age_s = divmod(max(0, snap_age), 60)
+
+    # Five-hour reset-in text (h, m precision)
+    if cw.five_hour_resets_at:
+        fr_delta = cw.five_hour_resets_at - snap.generated_at
+        fr_hr = max(0, int(fr_delta.total_seconds()) // 3600)
+        fr_mn = max(0, (int(fr_delta.total_seconds()) % 3600) // 60)
+        fr_str = f"resets in {fr_hr}h {fr_mn:02d}m"
+    else:
+        fr_str = ""
+
+    dpp_str = (
+        f"${cw.dollars_per_percent:.2f}"
+        if cw.dollars_per_percent is not None else "—"
+    )
+    lines = [
+        "",
+        f" Used   {{{used_cls}}}{bar_fill}{{/}} {{{used_cls}.b}}{cw.used_pct:>5.1f}%{{/}}",
+        "",
+        f" 5-hour {{bar.accent}}{five_bar}{{/}} {{bright}}{int(five):>3d}%{{/}}",
+        f"        {{dim}}{fr_str}{{/}}" if fr_str else "",
+        "",
+        f" {{dim}}Spent{{/}}    {{bright}}${cw.spent_usd:.2f}{{/}}        "
+        f"{{dim}}$/1%{{/}}  {{bright}}{dpp_str}{{/}}",
+        f" {{dim}}Reset{{/}}    {{bright}}{format_display_dt(cw.week_end_at, runtime.display_tz, fmt='%b %d %H:%M', suffix=True)}{{/}}  "
+        f"{{dim}}(in {reset_days}d {reset_hrs}h){{/}}",
+        "",
+        f" {{faint}}· last snapshot: {snap_age_m}m {snap_age_s:02d}s ago{{/}}",
+    ]
+    # Freshness chip (Task C6 / spec §3.4). Hidden when label is None or
+    # 'fresh'; rendered dim for 'aging', warn (amber) for 'stale'. Mirrors
+    # the dashboard CurrentWeekPanel chip in dashboard/web/src/panels/.
+    if cw.freshness_label and cw.freshness_label != "fresh":
+        captured_hms = format_display_dt(
+            cw.latest_snapshot_at, runtime.display_tz,
+            fmt="%H:%M:%S", suffix=False,
+        )
+        chip_style = "warn" if cw.freshness_label == "stale" else "dim"
+        chip_age = cw.freshness_age if cw.freshness_age is not None else 0
+        lines.append(
+            f"  {{{chip_style}}}⏱ as of {captured_hms} · {chip_age}s ago{{/}}"
+        )
+    return lines
+
+
+def _tui_panel_current_week_hero(
+    snap: DataSnapshot,
+    runtime: RuntimeState,
+    width: int,
+) -> list[str]:
+    """Variant B hero meter for current week."""
+    cw = snap.current_week
+    if cw is None:
+        return ["", "  {dim}no data yet{/}", ""]
+    bar_w = max(10, width - 10)
+    used_cls = _tui_bar_color(cw.used_pct)
+    big_bar = _tui_bar_string(cw.used_pct, bar_w)
+    five_bar = _tui_bar_string(cw.five_hour_pct or 0.0, bar_w)
+    snap_age_min = int((snap.generated_at - cw.latest_snapshot_at).total_seconds()) // 60
+
+    if cw.five_hour_resets_at:
+        sec = int((cw.five_hour_resets_at - snap.generated_at).total_seconds())
+        fr_hr = max(0, sec) // 3600
+        fr_mn = (max(0, sec) % 3600) // 60
+        reset_suffix = f"   {{dim}}resets {fr_hr}h {fr_mn:02d}m{{/}}"
+    else:
+        reset_suffix = ""
+
+    if snap.last_sync_error:
+        health = "{warn}daemon error{/}"
+    elif snap.last_sync_at is None:
+        health = "{dim}sync paused{/}"
+    else:
+        health = "{dim}daemon healthy{/}"
+
+    return [
+        "",
+        "  {dim}WEEK USAGE{/}",
+        "",
+        f"     {{{used_cls}.b}}{cw.used_pct:.1f}%{{/}}  {{dim}}of allowance used{{/}}",
+        "",
+        f"  {{{used_cls}}}{big_bar}{{/}}",
+        f"  {{faint}}0%{' ' * (bar_w - 6)}100%{{/}}",
+        "",
+        f"  {{dim}}5-HOUR WINDOW{{/}}  {{bright}}{int(cw.five_hour_pct or 0)}%{{/}}{reset_suffix}",
+        f"  {{bar.accent}}{five_bar}{{/}}",
+        "",
+        f"  {{dim}}snapshot {snap_age_min}m ago{{/}} · {health}",
+        "",
+    ]
+
+
+_TUI_VERDICT_CLS = {
+    "GOOD": "ok", "WARN": "warn", "OVER": "bad", "LOW CONF": "warn",
+}
+_TUI_VERDICT_SHORT = {
+    "GOOD": "comfortable headroom",
+    "WARN": "on track, no slack",
+    "OVER": "throttle immediately",
+    "LOW CONF": "not enough data",
+}
+
+
+def _tui_verdict_of(forecast) -> str:
+    """Compute verdict name from a ForecastOutput. Matches design language."""
+    if forecast is None or getattr(forecast.inputs, "confidence", "high") == "low":
+        return "LOW CONF"
+    high = forecast.final_percent_high
+    if high >= 100:
+        return "OVER"
+    if high >= 90:
+        return "WARN"
+    return "GOOD"
+
+
+def _tui_panel_forecast(
+    snap: DataSnapshot,
+    runtime: RuntimeState,
+    width: int,
+) -> list[str]:
+    """Variant A forecast panel body."""
+    fc = snap.forecast
+    if fc is None:
+        return [
+            "",
+            "  {badge.warn} [ LOW CONF ] {/} {dim}no current-week data{/}",
+            "",
+            "  {dim}run record-usage first{/}",
+            "",
+        ]
+    verdict = _tui_verdict_of(fc)
+    vcls = _TUI_VERDICT_CLS[verdict]
+    vmsg = _TUI_VERDICT_SHORT[verdict]
+
+    bar_w = max(6, width - 36)
+
+    def bar_tagged(val: float) -> str:
+        b = _tui_bar_string(min(val, 100), bar_w)
+        cls = _tui_bar_color(val)
+        return f"{{{cls}}}{b}{{/}}"
+
+    # Compute the two projection values DIRECTLY from the rate methods,
+    # not from final_low/final_high which are min/max aggregates and
+    # swap labels when the recent-24h rate is lower than week-avg.
+    p_now = fc.inputs.p_now
+    remaining = fc.inputs.remaining_hours
+    wa = int(round(p_now + fc.r_avg * remaining))
+    rc = wa if fc.r_recent is None else int(round(p_now + fc.r_recent * remaining))
+    # Budget table row values
+    b100 = next((r for r in fc.budgets if r.target_percent == 100), None)
+    b90 = next((r for r in fc.budgets if r.target_percent == 90), None)
+    b100_str = f"${b100.dollars_per_day:.2f}/day" if b100 and b100.dollars_per_day is not None else "—"
+    b90_str  = f"${b90.dollars_per_day:.2f}/day"  if b90  and b90.dollars_per_day  is not None else "—"
+    conf = "low" if verdict == "LOW CONF" else "high"
+
+    return [
+        "",
+        f"  {{badge.{vcls}}} [ {verdict} ] {{/}} {{dim}}{vmsg}{{/}}",
+        "",
+        f" {{dim}}Projection by week-avg{{/}}    {bar_tagged(wa)} {{bright}}{wa:>3d}%{{/}}",
+        f" {{dim}}Projection by recent 24h{{/}}  {bar_tagged(rc)} {{bright}}{rc:>3d}%{{/}}",
+        "",
+        f" {{dim}}Budget to stay ≤100%{{/}}   {{bright}}{b100_str}{{/}}",
+        f" {{dim}}Budget to stay  ≤90%{{/}}   {{bright}}{b90_str}{{/}}",
+        "",
+        f" {{faint}}confidence: {conf} · based on 7-day rate{{/}}",
+    ]
+
+
+def _tui_panel_trend(
+    snap: DataSnapshot,
+    runtime: RuntimeState,
+    width: int,
+    *,
+    compact: bool = False,
+) -> list[str]:
+    """Variant A trend panel: 8-row table + inline sparkline row.
+
+    When ``compact=True``, the leading blank, the pre-sparkline blank, and
+    the trailing blank are skipped (3 rows recovered) so callers with tight
+    vertical budgets can use the panel without the default padding.
+    """
+    rows = snap.trend
+    if not rows:
+        return ["", "  {dim}no trend data yet{/}", ""]
+    lines: list[str] = []
+    if not compact:
+        lines.append("")
+    lines.append(" {dim.b}Week      Used%    $/1%    Δ{/}")
+    lines.append(" {faint}────────── ───── ──────── ──────{/}")
+    for r in rows:
+        marker = "{accent}▶{/}" if r.is_current else " "
+        if r.used_pct is None:
+            used_cls = "dim"
+            used_fmt = "   — "  # 6 cols, matches "{:>5.1f}%" width
+        else:
+            used_cls = _tui_bar_color(r.used_pct)
+            used_fmt = f"{r.used_pct:>5.1f}%"
+        rate_str = (
+            f"${r.dollars_per_percent:.2f}"
+            if r.dollars_per_percent is not None else "   —"
+        )
+        if r.delta_dpp is None:
+            delta_str = "  —  "
+            delta_cls = "dim"
+        else:
+            sign = "+" if r.delta_dpp >= 0 else ""
+            delta_str = f"{sign}{r.delta_dpp:.2f}"
+            delta_cls = ("dim" if abs(r.delta_dpp) < 0.02
+                         else ("warn" if r.delta_dpp > 0 else "ok"))
+        lines.append(
+            f" {marker} {{bright}}{r.week_label:<9}{{/}}  "
+            f"{{{used_cls}}}{used_fmt}{{/}}   "
+            f"{{bright}}{rate_str:<6}{{/}}  {{{delta_cls}}}{delta_str:<5}{{/}}"
+        )
+    # Sparkline row
+    if not compact:
+        lines.append("")
+    heights = [r.spark_height for r in rows]
+    spark = _tui_sparkline_inline(heights)
+    lines.append(f"   {{dim}}spark $/1%{{/}}   {{accent.b}}{spark}{{/}}")
+    if not compact:
+        lines.append("")
+    return lines
+
+
+def _tui_session_model_cls(model: str) -> str:
+    """Map primary model name to a color class for the Model column."""
+    m = (model or "").lower()
+    if m.startswith("opus"):
+        return "magenta"
+    if m.startswith("haiku"):
+        return "blue"
+    return "bright"
+
+
+def _tui_format_started(
+    ts: dt.datetime,
+    now: dt.datetime,
+    tz: "ZoneInfo | None",
+) -> str:
+    """Today -> 'HH:MM:SS', else 'Mon DD HH:MM'.
+
+    F4 fix: takes a resolved ``ZoneInfo | None`` (as carried on
+    ``RuntimeState.display_tz``) instead of the legacy "utc" / "local"
+    string token. Previously, an explicit IANA zone like
+    ``America/New_York`` reached this helper as a string, took the else
+    branch, and rendered the raw UTC clock — so session rows displayed
+    UTC even when reset/session-detail fields used the resolved zone.
+    """
+    # internal fallback: host-local intentional — picks the calendar bucket;
+    # the actual rendered string flows through `format_display_dt`.
+    disp = ts.astimezone(tz) if tz is not None else ts.astimezone()
+    today = now.astimezone(disp.tzinfo).date()
+    if disp.date() == today:
+        return format_display_dt(ts, tz, fmt="%H:%M:%S", suffix=False)
+    return format_display_dt(ts, tz, fmt="%b %d %H:%M", suffix=False)
+
+
+def _tui_format_dur(minutes: float) -> str:
+    """Human-friendly duration: '42m' or '3h 07m'."""
+    if minutes < 60:
+        return f"{int(minutes)}m"
+    h = int(minutes // 60)
+    m = int(minutes % 60)
+    return f"{h}h {m:02d}m"
+
+
+# v2 sort: cycle order + direction (spec §3.1).
+_TUI_SORT_KEYS = ("last-activity", "cost", "duration", "model", "project")
+_TUI_SORT_ASC = frozenset({"model", "project"})  # ascending; rest descending
+
+
+def _tui_sort_sessions(sessions: list[TuiSessionRow], key: str) -> list[TuiSessionRow]:
+    """Return a new list sorted per spec §3.1.
+
+    Default key 'last-activity' is pass-through — preserves the order
+    `_aggregate_claude_sessions` already produces (last_activity desc).
+    Other keys: hard-coded direction by type (numeric/recency desc;
+    text asc) with a stable secondary on last-activity desc.
+    """
+    if not sessions:
+        return []
+    if key == "last-activity":
+        return list(sessions)  # already sorted by aggregator — pass-through
+    if key == "cost":
+        primary = lambda s: -s.cost_usd
+    elif key == "duration":
+        primary = lambda s: -s.duration_minutes
+    elif key == "model":
+        primary = lambda s: s.model_primary.lower()
+    elif key == "project":
+        primary = lambda s: s.project_label.lower()
+    else:
+        return list(sessions)
+    return sorted(sessions, key=lambda s: (primary(s), -s.started_at.timestamp()))
+
+
+def _tui_next_sort_key(current: str) -> str:
+    """Cycle to the next key. Wraps."""
+    try:
+        idx = _TUI_SORT_KEYS.index(current)
+    except ValueError:
+        return _TUI_SORT_KEYS[0]
+    return _TUI_SORT_KEYS[(idx + 1) % len(_TUI_SORT_KEYS)]
+
+
+def _tui_apply_session_filter(sessions, active_filter):
+    """Narrow sessions by a filter substring (project_label|model_primary).
+    Returns `sessions` unchanged when `active_filter` is None/empty. Mirrors
+    the filter logic in `_tui_panel_sessions` so live match counts in the
+    input prompt reflect the post-filter navigable list."""
+    if not active_filter:
+        return sessions
+    af_lower = active_filter.lower()
+    return [
+        s for s in sessions
+        if af_lower in s.project_label.lower()
+        or af_lower in s.model_primary.lower()
+    ]
+
+
+def _tui_sessions_title(runtime: RuntimeState, *, narrow: bool) -> str:
+    """Build the Sessions panel title with sort indicator and filter chip.
+
+    Spec §3.1 (sort indicator) + §3.2 (filter chip). Narrow bucket
+    abbreviates the sort key and chip per spec §5.1.
+
+    Returns a tagged string (uses {name}…{/} markup); the caller passes
+    it straight to `_tui_tagged_box_lines(title=…)` which materializes
+    the tags via `_tui_colortag` downstream.
+    """
+    arrow = "↑" if runtime.sort_key in _TUI_SORT_ASC else "↓"
+    if narrow:
+        sort_part = f"{{dim}} · {runtime.sort_key}{arrow}{{/}}"
+    else:
+        sort_part = f"{{dim}} · sort: {runtime.sort_key} {arrow}{{/}}"
+    chip_part = ""
+    if runtime.filter_term is not None:
+        if narrow:
+            shown = _tui_escape_tags(runtime.filter_term[:8])
+            chip_part = f" {{chip}} ▼{shown} {{/}}"
+        else:
+            shown = _tui_escape_tags(runtime.filter_term)
+            chip_part = f" {{chip}} filter: {shown} {{/}}"
+    return f"{{focused.b}}Recent Sessions{{/}}{sort_part}{chip_part}"
+
+
+def _tui_panel_sessions(
+    snap: DataSnapshot,
+    runtime: RuntimeState,
+    width: int,
+    *,
+    rows_visible: int,
+    show_project_col: bool,
+) -> list[str]:
+    """Variant A + B sessions panel body.
+
+    Caller drives layout:
+    - Variant A (right half of the 2x2) passes rightW as `width` and
+      `show_project_col=False` or True depending on space.
+    - Variant B (full-width) passes the terminal width as `width` and
+      typically `show_project_col=True`.
+    """
+    sessions = _tui_sort_sessions(snap.sessions, runtime.sort_key)
+
+    # v2: apply filter (spec §3.2). Use input_buffer as the live preview when
+    # in filter input mode (incremental narrowing); otherwise use the
+    # committed filter_term.
+    active_filter: str | None
+    if runtime.input_mode == "filter":
+        active_filter = runtime.input_buffer or None
+    else:
+        active_filter = runtime.filter_term
+    if active_filter:
+        af_lower = active_filter.lower()
+        sessions = [
+            s for s in sessions
+            if af_lower in s.project_label.lower()
+            or af_lower in s.model_primary.lower()
+        ]
+
+    # Spec §3.2 empty-narrow result.
+    if active_filter and not sessions:
+        empty_lines: list[str] = [
+            "",
+            "",
+            f"  {{dim}}no sessions match \"{_tui_escape_tags(active_filter)}\"  · f to edit{{/}}",
+            "",
+        ]
+        # Pad to expected rows_visible (header+ruler+rows+trailing chrome).
+        while len(empty_lines) < rows_visible + 5:
+            empty_lines.append("")
+        empty_lines.append("")
+        empty_lines.append(
+            f" {{dim}}↑↓ scroll · 0 of {len(snap.sessions)} match · 0 below{{/}}"
+        )
+        return empty_lines
+
+    # v2: compute search matches (spec §3.3). Active term is input_buffer
+    # while typing, search_term once confirmed, None otherwise.
+    if runtime.input_mode == "search":
+        active_search = runtime.input_buffer or None
+        auto_jump = True  # live-type: reset scroll to first match each tick
+    else:
+        active_search = runtime.search_term
+        auto_jump = False  # confirmed: preserve scroll (lets n/N work)
+    match_indices: list[int] = []
+    if active_search:
+        as_lower = active_search.lower()
+        for i, s in enumerate(sessions):
+            haystack = (
+                s.project_label.lower() + "|"
+                + s.model_primary.lower() + "|"
+                + _tui_format_started(s.started_at, snap.generated_at, runtime.display_tz).lower()
+            )
+            if as_lower in haystack:
+                match_indices.append(i)
+        # Live jump only while typing — n/N drives scroll after confirm.
+        if match_indices and auto_jump:
+            runtime.session_scroll = match_indices[0]
+        # Persist for n/N when in confirmed state.
+        runtime.search_matches = match_indices
+        if runtime.search_index >= len(match_indices):
+            runtime.search_index = 0
+
+    def _hl(text: str, term: str | None) -> str:
+        if not term:
+            return text
+        # Search and slice on UNESCAPED text (so positions are correct), but
+        # emit ESCAPED segments so any user-source `{` `}` chars survive
+        # the colortag pipeline literally.
+        idx = text.lower().find(term.lower())
+        if idx < 0:
+            return text
+        before = _tui_escape_tags(text[:idx])
+        match  = _tui_escape_tags(text[idx:idx + len(term)])
+        after  = _tui_escape_tags(text[idx + len(term):])
+        return before + "{match}" + match + "{/}" + after
+
+    interior = width - 2
+
+    # Column widths derived from the design's 120-col and 100-col tables.
+    # Cost column is 8 at the wide bucket so $1000+ session values don't
+    # overflow (e.g., "$1234.56" is 8 chars). Medium/narrow buckets keep 7/6
+    # — the layouts are already tight and four-digit costs are uncommon.
+    if interior >= 70:
+        c_start, c_dur, c_model, c_cost = 14, 6, 11, 8
+    elif interior >= 55:
+        c_start, c_dur, c_model, c_cost = 12, 6, 10, 7
+    else:
+        c_start, c_dur, c_model, c_cost = 8, 5, 10, 6
+        show_project_col = False
+
+    fixed = 8 + c_start + c_dur + c_model + c_cost
+    last_avail = interior - fixed
+    if show_project_col and last_avail >= 10:
+        c_last = last_avail
+        last_title = "Project"
+        use_project = True
+    else:
+        c_last = min(5, max(0, last_avail))
+        last_title = "Cache"
+        use_project = False
+
+    def _truncpad(s: str, n: int) -> str:
+        if n <= 0:
+            return ""
+        if len(s) > n:
+            return s[: n - 1] + "…"
+        return s + " " * (n - len(s))
+
+    lines: list[str] = [""]
+    # Header row
+    lines.append(
+        f"   {{dim.b}}"
+        f"{_truncpad('Started', c_start)} "
+        f"{_truncpad('Dur', c_dur)} "
+        f"{_truncpad('Model', c_model)} "
+        f"{('Cost').rjust(c_cost)} "
+        f"{_truncpad(last_title, c_last)}"
+        f"{{/}}"
+    )
+    # Ruler
+    lines.append(
+        f"   {{faint}}"
+        f"{'─' * c_start} {'─' * c_dur} {'─' * c_model} "
+        f"{'─' * c_cost} {'─' * c_last}"
+        f"{{/}}"
+    )
+
+    # Clamp scroll. Cap at len-1 (not len - rows_visible) so any session —
+    # including search matches landing in the last rows_visible-1 positions —
+    # can become the topmost-visible row. Spec §3.3 (Live jump) requires the
+    # matched row to reach topmost; with the tighter clamp, selection (=
+    # topmost per the "is_selected = i == 0" convention) would diverge from
+    # the highlighted match and Enter would open the wrong session detail.
+    # The existing padding loop below fills blank rows when the slice is short.
+    max_scroll = max(0, len(sessions) - 1)
+    scroll = min(runtime.session_scroll, max_scroll)
+    runtime.session_scroll = scroll  # write clamped back
+
+    visible = sessions[scroll: scroll + rows_visible]
+    for i, s in enumerate(visible):
+        is_selected = (i == 0)  # topmost = selected (design convention)
+        start_s = _truncpad(_tui_format_started(s.started_at, snap.generated_at, runtime.display_tz), c_start)
+        dur_s   = _truncpad(_tui_format_dur(s.duration_minutes), c_dur)
+        model_s = _truncpad(s.model_primary, c_model)
+        cost_s  = f"${s.cost_usd:.2f}".rjust(c_cost)
+        if use_project:
+            last_s = _truncpad(s.project_label, c_last)
+        else:
+            last_s = (f"{int(s.cache_hit_pct)}%" if s.cache_hit_pct is not None else "—").rjust(c_last)
+        model_cls = _tui_session_model_cls(s.model_primary)
+        cache_cls = "ok" if (s.cache_hit_pct or 0) >= 70 else ("warn" if (s.cache_hit_pct or 0) >= 50 else "dim")
+
+        # v2: search highlight (spec §3.3). Wraps matched substring in
+        # {match} tags; nests safely inside the row's outer style.
+        model_h = _hl(model_s, active_search)
+        last_h = _hl(last_s, active_search) if use_project else last_s
+        start_h = _hl(start_s, active_search)
+
+        if is_selected:
+            body = (
+                f"▸ {start_h} {dur_s} {model_h} {cost_s} {last_h}"
+            )
+            lines.append(f" {{focused.b}}{body}{{/}}")
+        elif use_project:
+            lines.append(
+                f"   {{bright}}{start_h}{{/}} {{dim}}{dur_s}{{/}} "
+                f"{{{model_cls}}}{model_h}{{/}} {{bright}}{cost_s}{{/}} "
+                f"{{dim}}{last_h}{{/}}"
+            )
+        else:
+            lines.append(
+                f"   {{bright}}{start_h}{{/}} {{dim}}{dur_s}{{/}} "
+                f"{{{model_cls}}}{model_h}{{/}} {{bright}}{cost_s}{{/}} "
+                f"{{{cache_cls}}}{last_s}{{/}}"
+            )
+    # Pad to rows_visible
+    while len(lines) - 3 < rows_visible:
+        lines.append("")
+
+    below = max(0, len(sessions) - scroll - rows_visible)
+    lines.append("")
+    lines.append(f" {{dim}}↑↓ scroll · {below} below{{/}}")
+    return lines
+
+
+def _tui_header_strip_a(
+    snap: DataSnapshot, runtime: RuntimeState, width: int,
+) -> list[str]:
+    """Variant A header strip: top rule + summary line + bottom rule.
+
+    Appends a one-line error banner below the bottom rule when
+    snap.last_sync_error is set.
+    """
+    import time
+    cw = snap.current_week
+    fc = snap.forecast
+    verdict = _tui_verdict_of(fc) if fc else "LOW CONF"
+    vcls = _TUI_VERDICT_CLS[verdict]
+    sync_age = 0
+    if snap.last_sync_at is not None:
+        sync_age = int(time.monotonic() - snap.last_sync_at)
+    sync_txt = f"synced {sync_age}s ago" if snap.last_sync_at is not None else "synced —"
+    err = snap.last_sync_error
+    if cw is None:
+        hdr = (
+            f"{{bright.b}}Week — {{/}} {{faint}}│{{/}} "
+            f"{{dim}}no data yet — run record-usage first{{/}}"
+        )
+    else:
+        used_cls = _tui_bar_color(cw.used_pct)
+        dpp_str = (
+            f"${cw.dollars_per_percent:.2f}"
+            if cw.dollars_per_percent is not None else "—"
+        )
+        fcst_pct = "—"
+        if fc:
+            # Use the measure that drives the verdict (_tui_verdict_of keys
+            # on final_percent_high). Using low here would display e.g.
+            # "Fcst 74% WARN" where the WARN comes from a >=90% high
+            # projection, understating risk in the most glanceable line.
+            fcst_pct = f"{int(round(fc.final_percent_high))}%"
+        hdr = (
+            f"{{bright.b}}Week {format_display_dt(cw.week_start_at, runtime.display_tz, fmt='%b %d', suffix=False)}–{format_display_dt(cw.week_end_at, runtime.display_tz, fmt='%b %d', suffix=False)}{{/}} "
+            f"{{faint}}│{{/}} Used {{{used_cls}.b}}{cw.used_pct:.1f}%{{/}} "
+            f"{{dim}}(5h {int(cw.five_hour_pct or 0)}%){{/}} {{faint}}│{{/}} "
+            f"$/1% {{bright}}{dpp_str}{{/}} {{faint}}│{{/}} "
+            f"Fcst {{{vcls}.b}}{fcst_pct}{{/}} {{{vcls}.b}}{verdict}{{/}} {{faint}}│{{/}} "
+            f"{{dim.pulse}}● {sync_txt}{{/}}"
+        )
+    # Top/bottom rules framing the header.
+    return [
+        "{faint}" + ("═" * width) + "{/}",
+        " " + hdr,
+        "{faint}" + ("═" * width) + "{/}",
+        *(["{warn}⚠ sync failed: " + err + "{/}"] if err else []),
+    ]
+
+
+def _tui_footer_keys(width: int) -> list[str]:
+    """Variant A footer: top rule + keys legend."""
+    return [
+        "{faint}" + ("═" * width) + "{/}",
+        (" {bright}Tab{/} {dim}focus{/}  "
+         "{bright}↑↓{/} {dim}scroll{/}  "
+         "{bright}r{/} {dim}refresh{/}  "
+         "{bright}s{/} {dim}sort{/}  "
+         "{bright}f{/} {dim}filter{/}  "
+         "{bright}/{/} {dim}search{/}  "
+         "{bright}Enter{/} {dim}open{/}  "
+         "{bright}v{/} {dim}variant{/}  "
+         "{bright}?{/} {dim}help{/}  "
+         "{bright}q{/} {dim}quit{/}"),
+    ]
+
+
+def _tui_render_input_prompt(
+    runtime: RuntimeState, width: int, *, match_count: int | None = None,
+) -> list[str]:
+    """Render the bottom-row input prompt. Replaces the keys-legend row
+    while runtime.input_mode is set. Spec §3.2 (filter), §3.3 (search).
+    """
+    buf = runtime.input_buffer
+    # Truncate displayed buffer if it would overflow.
+    max_buf = max(10, width - 60)
+    shown = buf if len(buf) <= max_buf else buf[-max_buf:]
+    # Escape user input so a stray `{` or `}` doesn't get parsed as a
+    # style tag by _tui_colortag (would crash the live render loop).
+    shown = _tui_escape_tags(shown)
+    if runtime.input_mode == "filter":
+        prefix = "filter (project|model)"
+        contract = "enter apply · esc cancel"
+    else:  # 'search'
+        prefix = "search"
+        contract = "enter confirm · esc cancel · n/N next/prev"
+    match_suffix = ""
+    if match_count is not None:
+        cls = "bad" if match_count == 0 else "dim"
+        match_suffix = f" {{{cls}}}· {match_count} matches{{/}}"
+    body = (f" {{prompt}}{prefix}:{{/}} {{bright}}{shown}{{/}}{{caret}} {{/}}"
+            f"{match_suffix}     {{faint}}{contract}{{/}}")
+    return [
+        "{faint}" + ("═" * width) + "{/}",
+        body,
+    ]
+
+
+# Tag matcher used to strip color tags for plain-text width math. Matches
+# both opening tags ({name} or {name.mod}) and closing tags ({/}).
+_TUI_TAG_RE = re.compile(r"\{(?:/|[a-zA-Z.]+)\}")
+
+
+def _tui_strip_tags(s: str) -> str:
+    """Return ``s`` with all color-tag markup removed (for width math)."""
+    return _TUI_TAG_RE.sub("", s)
+
+
+def _tui_tagged_box_lines(
+    *,
+    width: int,
+    body_tagged: list[str],
+    title: str | None,
+    pin: str | None,
+    border_style: str = "faint",
+) -> list[str]:
+    """Return a list of tagged strings forming a double-line box.
+
+    Body lines may contain color tags. Width math strips tags before padding
+    so that color markup does not inflate the visible length. ``border_style``
+    is a theme style name applied to all border glyphs.
+
+    When a body line's plain length exceeds the interior width it is truncated
+    with ``{/}`` appended as a safety net — callers should size their content
+    to avoid this branch.
+    """
+    H, V = _TUI_BOX["h"], _TUI_BOX["v"]
+    TL, TR, BL, BR = _TUI_BOX["tl"], _TUI_BOX["tr"], _TUI_BOX["bl"], _TUI_BOX["br"]
+    interior = width - 2
+
+    def _wrap_border(s: str) -> str:
+        return f"{{{border_style}}}{s}{{/}}"
+
+    def _top() -> str:
+        if title is None and pin is None:
+            return _wrap_border(TL + H * interior + TR)
+        t_seg = f" {title} " if title else ""
+        p_seg = f" {pin} " if pin else ""
+        if title and pin:
+            fill = width - 4 - len(_tui_strip_tags(t_seg)) - len(_tui_strip_tags(p_seg))
+            if fill >= 1:
+                return (_wrap_border(TL + H) + t_seg
+                        + _wrap_border(H * fill) + p_seg + _wrap_border(H + TR))
+        if title:
+            fill = width - 3 - len(_tui_strip_tags(t_seg))
+            if fill >= 1:
+                return _wrap_border(TL + H) + t_seg + _wrap_border(H * fill + TR)
+        return _wrap_border(TL + H * interior + TR)
+
+    lines: list[str] = [_top()]
+    for line in body_tagged:
+        plain = _tui_strip_tags(line)
+        if len(plain) > interior:
+            # On overflow, drop color markup entirely — partial-tag truncation
+            # would produce malformed tokens that crash _tui_colortag. Plain-
+            # text is always safe, and callers are expected to size content
+            # to fit (this branch is a safety net only).
+            line = plain[: max(0, interior - 1)] + "…"
+            plain = line
+        pad = interior - len(plain)
+        lines.append(_wrap_border(V) + line + " " * pad + _wrap_border(V))
+    lines.append(_wrap_border(BL + H * interior + BR))
+    return lines
+
+
+def _tui_lines_to_text(lines: list[str]):
+    """Join a list of tagged strings into a single rich.text.Text blob.
+
+    Each line is passed through ``_tui_colortag`` to materialize the style
+    tags; adjacent lines are separated by a literal ``"\\n"``.
+    """
+    from rich.text import Text
+    out = Text()
+    for i, l in enumerate(lines):
+        if i:
+            out.append("\n")
+        out.append(_tui_colortag(l))
+    return out
+
+
+def _tui_render_variant_a(
+    snap: DataSnapshot, runtime: RuntimeState,
+    width: int, height: int, bucket: str,
+    *, overlay_panel: Panel | None = None,
+) -> Layout:
+    """Return a ``rich.layout.Layout`` for the whole Variant A frame.
+
+    Assembles the 2x2 grid (Current Week | Forecast / Trend | Sessions)
+    with header and footer strips.
+
+    The returned Layout has real sub-regions:
+      root (split_column):
+        - header     (size = len(header_lines))
+        - warn?      (size = 1, only when bucket == "narrow")
+        - top_row    (split_row current_week + forecast)
+        - sep        (size = 1)
+        - bottom_row (split_row trend + sessions)
+        - footer     (size = len(footer_lines))
+
+    When ``overlay_panel`` is provided, the body regions (top_row/sep/bottom_row)
+    are replaced with a single centered Align(overlay_panel) so the help
+    overlay or v2 detail modal appears in the dashboard's body area.
+    Header and footer remain visible. This is the body-region-swap overlay
+    composition (Fallback A).
+    """
+    from rich.layout import Layout
+    from rich.align import Align
+
+    left_w = width // 2
+    right_w = width - left_w
+
+    header = _tui_header_strip_a(snap, runtime, width)
+    if runtime.input_mode is not None:
+        match_count = None
+        if runtime.input_mode == "filter":
+            match_count = sum(
+                1 for s in snap.sessions
+                if (runtime.input_buffer.lower() in s.project_label.lower()
+                    or runtime.input_buffer.lower() in s.model_primary.lower())
+            ) if runtime.input_buffer else None
+        elif runtime.input_mode == "search":
+            if runtime.input_buffer:
+                needle = runtime.input_buffer.lower()
+                # Count against the post-filter list: confirmed search matches
+                # are computed against the already-filtered sessions in
+                # _tui_panel_sessions, so the live count must use the same set
+                # or the prompt overstates what n/N can reach.
+                pool = _tui_apply_session_filter(snap.sessions, runtime.filter_term)
+                count = 0
+                for s in pool:
+                    hay = (
+                        s.project_label.lower() + "|"
+                        + s.model_primary.lower() + "|"
+                        + _tui_format_started(s.started_at, snap.generated_at, runtime.display_tz).lower()
+                    )
+                    if needle in hay:
+                        count += 1
+                match_count = count
+            else:
+                match_count = None
+        footer = _tui_render_input_prompt(runtime, width, match_count=match_count)
+    else:
+        footer = _tui_footer_keys(width)
+
+    warn_line = (
+        "{warn}⚠ narrow terminal — some columns hidden{/}"
+        if bucket == "narrow" else None
+    )
+
+    # Compute the body region (between header and footer).
+    # When a narrow-warning line is present it occupies one extra row that
+    # would otherwise belong to the body, matching the legacy behavior of
+    # inserting the warn line at frame-row index 3.
+    warn_rows = 1 if warn_line is not None else 0
+    body_height = max(
+        10,
+        height - len(header) - warn_rows - len(footer) - 1,
+    )
+    top_h = body_height // 2
+    bot_h = body_height - top_h
+
+    # TOP: current week (left) | forecast (right)
+    cw_body = _tui_panel_current_week(
+        snap, runtime, left_w, focused=runtime.focus_index == 0
+    )
+    fc_body = _tui_panel_forecast(snap, runtime, right_w)
+    cw_box = _tui_tagged_box_lines(
+        width=left_w, body_tagged=cw_body,
+        title="{accent.b}Current Week{/}", pin="{dim}[1]{/}",
+        border_style=("focused" if runtime.focus_index == 0 else "faint"),
+    )
+    fc_box = _tui_tagged_box_lines(
+        width=right_w, body_tagged=fc_body,
+        title="{accent.b}Forecast{/}", pin="{dim}[2]{/}",
+        border_style=("focused" if runtime.focus_index == 1 else "faint"),
+    )
+    # Pad shorter box with blank interior rows up to the taller one.
+    maxlen = max(len(cw_box), len(fc_box))
+
+    def _pad_box(lines: list[str], w: int, style: str) -> list[str]:
+        while len(lines) < maxlen:
+            lines.insert(-1, f"{{{style}}}║{{/}}{' ' * (w - 2)}{{{style}}}║{{/}}")
+        return lines
+
+    cw_box = _pad_box(cw_box, left_w, "focused" if runtime.focus_index == 0 else "faint")
+    fc_box = _pad_box(fc_box, right_w, "focused" if runtime.focus_index == 1 else "faint")
+
+    # BOTTOM: trend (left) | sessions (right). Use compact trend when the
+    # cell height is tight (saves 3 rows of padding).
+    trend_compact = (bot_h - 2) < 14
+    trend_body = _tui_panel_trend(snap, runtime, left_w, compact=trend_compact)
+    # Sessions chrome: 1 leading blank + 1 header + 1 ruler + 1 trailing blank
+    # + 1 scroll footer = 5 rows around the data rows. Interior = bot_h - 2.
+    rows_visible = max(3, bot_h - 2 - 5)
+    show_proj = bucket == "wide"
+    sess_body = _tui_panel_sessions(
+        snap, runtime, right_w,
+        rows_visible=rows_visible,
+        show_project_col=show_proj,
+    )
+    trend_box = _tui_tagged_box_lines(
+        width=left_w, body_tagged=trend_body,
+        title="{accent.b}$/1% Trend{/} {dim}· 8 weeks{/}", pin="{dim}[3]{/}",
+        border_style=("focused" if runtime.focus_index == 2 else "faint"),
+    )
+    sess_box = _tui_tagged_box_lines(
+        width=right_w, body_tagged=sess_body,
+        # Variant A puts Sessions in a half-pane; compact bucket needs narrow form
+        # to keep the `focus` pin visible alongside the sort indicator. When a
+        # filter chip is active, use narrow form even at the wide bucket because
+        # the chip pushes the wide form past the half-pane width.
+        title=_tui_sessions_title(
+            runtime,
+            narrow=(bucket in ("narrow", "compact") or runtime.filter_term is not None),
+        ),
+        pin=("{focused}focus{/}" if runtime.focus_index == 3 else "{dim}[4]{/}"),
+        border_style=("focused" if runtime.focus_index == 3 else "faint"),
+    )
+    maxlen2 = max(len(trend_box), len(sess_box))
+
+    def _pad_box2(lines: list[str], w: int, style: str) -> list[str]:
+        while len(lines) < maxlen2:
+            lines.insert(-1, f"{{{style}}}║{{/}}{' ' * (w - 2)}{{{style}}}║{{/}}")
+        return lines
+
+    trend_box = _pad_box2(trend_box, left_w, "focused" if runtime.focus_index == 2 else "faint")
+    sess_box = _pad_box2(sess_box, right_w, "focused" if runtime.focus_index == 3 else "faint")
+
+    # Build per-region Text blobs. Each row band renders as a single Text
+    # so intra-band line breaks are not padded to the full width by
+    # Layout's line-pad; only the region-tail gets padded.
+    header_text = _tui_lines_to_text(header)
+    footer_text = _tui_lines_to_text(footer)
+    cw_text = _tui_lines_to_text(cw_box)
+    fc_text = _tui_lines_to_text(fc_box)
+    trend_text = _tui_lines_to_text(trend_box)
+    sess_text = _tui_lines_to_text(sess_box)
+
+    root = Layout()
+    regions: list[Layout] = [Layout(name="header", size=len(header))]
+    if warn_line is not None:
+        regions.append(Layout(name="warn", size=1))
+
+    if overlay_panel is not None:
+        # Fallback A: collapse the body bands into a single region containing
+        # a vertically-centered overlay Panel. Preserve header + footer.
+        body_rows = maxlen + 1 + maxlen2
+        regions.append(Layout(name="body", size=body_rows))
+        regions.append(Layout(name="footer", size=len(footer)))
+        root.split_column(*regions)
+        root["header"].update(header_text)
+        if warn_line is not None:
+            root["warn"].update(_tui_lines_to_text([warn_line]))
+        root["body"].update(Align.center(overlay_panel, vertical="middle"))
+        root["footer"].update(footer_text)
+        root._tui_natural_height = sum(r.size or 0 for r in regions)
+        return root
+
+    regions.extend([
+        Layout(name="top", size=maxlen),
+        Layout(name="sep", size=1),
+        Layout(name="bot", size=maxlen2),
+        Layout(name="footer", size=len(footer)),
+    ])
+    root.split_column(*regions)
+
+    root["header"].update(header_text)
+    if warn_line is not None:
+        root["warn"].update(_tui_lines_to_text([warn_line]))
+
+    root["top"].split_row(
+        Layout(name="cw", size=left_w),
+        Layout(name="fc", size=right_w),
+    )
+    root["top"]["cw"].update(cw_text)
+    root["top"]["fc"].update(fc_text)
+
+    # The separator is intentionally a single blank row. The leading rich
+    # Text of a single empty string becomes a zero-column line; Layout pads
+    # it to the full width (that padding is the trailing-whitespace drift
+    # accepted by the scoped-relaxation protocol).
+    root["sep"].update(_tui_lines_to_text([""]))
+
+    root["bot"].split_row(
+        Layout(name="trend", size=left_w),
+        Layout(name="sess", size=right_w),
+    )
+    root["bot"]["trend"].update(trend_text)
+    root["bot"]["sess"].update(sess_text)
+    root["footer"].update(footer_text)
+    # Stash the natural (content-filling) height so ``_tui_render_once``
+    # can recover the pre-refactor row count without padding to the full
+    # terminal. Live mode ignores this attribute — it lets Layout fill the
+    # actual terminal height, which is the desired TUI behavior.
+    root._tui_natural_height = sum(r.size or 0 for r in regions)
+    return root
+
+
+def _tui_render_variant_b(
+    snap: DataSnapshot, runtime: RuntimeState,
+    width: int, height: int, bucket: str,
+    *, overlay_panel: Panel | None = None,
+) -> Layout:
+    """Return a ``rich.layout.Layout`` for the whole Variant B frame.
+
+    Structure: ribbon -> subheader -> hero row (big meter + promoted
+    sparkline) -> forecast-budget strip -> full-width sessions -> footer.
+
+    Vertical bands as Layout regions:
+      ribbon (size=1), sub (size=1), rule (size=1),
+      warn? (size=1, narrow only), blank1 (size=1),
+      hero_row (split_row hero + trend, size=len(hero_box)),
+      blank2 (size=1), fc_strip (size=len(fc_strip_box)),
+      blank3 (size=1),
+      sessions (size=len(sess_box)), footer (size=len(footer_lines))
+
+    When ``overlay_panel`` is provided, the body bands collapse into one
+    ``body`` region filled with a centered Align(overlay_panel) so the
+    help overlay or v2 detail modal appears in the dashboard's body area
+    (Fallback A overlay). Ribbon / subheader / rule / footer remain visible.
+    """
+    from rich.layout import Layout
+    from rich.align import Align
+
+    # --- Ribbon ---------------------------------------------------------
+    verdict = _tui_verdict_of(snap.forecast) if snap.forecast else "LOW CONF"
+    vcls = _TUI_VERDICT_CLS[verdict]
+    # _TUI_VERDICT_CLS always maps to ok/warn/bad after the SSoT fix, so vcls
+    # is guaranteed to be a valid badge class here.
+    if snap.forecast:
+        # Compute projections directly from rate methods — final_low/final_high
+        # are min/max aggregates and swap labels when recent-24h rate is lower
+        # than week-average (mirrors the Variant A fix in commit 15b6fab).
+        p_now = snap.forecast.inputs.p_now
+        remaining = snap.forecast.inputs.remaining_hours
+        wa = int(round(p_now + snap.forecast.r_avg * remaining))
+        rc = wa if snap.forecast.r_recent is None else int(round(p_now + snap.forecast.r_recent * remaining))
+    else:
+        wa, rc = 0, 0
+    vmsg = _TUI_VERDICT_SHORT[verdict]
+    ribbon_text = f"  [ {verdict} ]   {vmsg}   ·   week-avg {wa}%   ·   recent-24h {rc}%"
+    ribbon_pad = max(0, width - len(ribbon_text))
+    ribbon = f"{{badge.{vcls}}}{ribbon_text}{' ' * ribbon_pad}{{/}}"
+
+    # --- Subheader ------------------------------------------------------
+    cw = snap.current_week
+    if cw:
+        import time as _t
+        sync_age = 0
+        if snap.last_sync_at is not None:
+            sync_age = int(_t.monotonic() - snap.last_sync_at)
+        sync_txt = f"synced {sync_age}s ago" if snap.last_sync_at is not None else "synced —"
+        # Pre-compute interpolated fragments so nothing inside the f-string
+        # uses a nested conditional format spec (which Python rejects).
+        dpp_str = (
+            f"${cw.dollars_per_percent:.2f}"
+            if cw.dollars_per_percent is not None else "—"
+        )
+        reset_delta = cw.week_end_at - snap.generated_at
+        reset_secs = max(0, int(reset_delta.total_seconds()))
+        reset_days = reset_secs // 86400
+        reset_hrs = (reset_secs % 86400) // 3600
+        sub = (
+            f" {{bright.b}}Week {format_display_dt(cw.week_start_at, runtime.display_tz, fmt='%b %d', suffix=False)}–"
+            f"{format_display_dt(cw.week_end_at, runtime.display_tz, fmt='%b %d', suffix=False)}{{/}}   "
+            f"{{dim}}${cw.spent_usd:.2f} spent · $/1% {dpp_str} · "
+            f"resets in {reset_days}d {reset_hrs}h{{/}}"
+            f"   ·   {{dim.pulse}}● {sync_txt}{{/}}"
+        )
+    else:
+        sub = " {dim}no current-week data yet{/}"
+
+    hero_left_w = 55 if bucket != "wide" else int(width * 0.56)
+    hero_right_w = width - hero_left_w
+
+    hero_body = _tui_panel_current_week_hero(snap, runtime, hero_left_w)
+
+    # Big sparkline on the right (promoted trend view).
+    heights = [r.spark_height for r in snap.trend] or [1]
+    big = _tui_sparkline_big(heights).split("\n")
+    cur_rate = (snap.trend[-1].dollars_per_percent if snap.trend else None)
+    cur_delta = (snap.trend[-1].delta_dpp if snap.trend else None)
+    rate_str = f"${cur_rate:.2f}" if cur_rate is not None else "—"
+    if cur_delta is None:
+        delta_str = "—"
+    else:
+        sign = "+" if cur_delta >= 0 else ""
+        delta_str = f"{sign}{cur_delta:.2f}"
+
+    trend_title_text = "this week" if bucket != "wide" else "$/1% OVER 8 WEEKS"
+    trend_body = [
+        "",
+        f"  {{dim}}{trend_title_text}{{/}}  "
+        f"{{bright.b}}{rate_str}{{/}}  {{ok}}{delta_str}{{/}}",
+        "",
+        f"  {{accent}}{big[0] if len(big) > 0 else ''}{{/}}",
+        f"  {{accent}}{big[1] if len(big) > 1 else ''}{{/}}",
+        f"  {{accent}}{big[2] if len(big) > 2 else ''}{{/}}",
+        f"  {{faint}}{'─' * min(hero_right_w - 4, 24)}{{/}}",
+        "",
+    ]
+
+    hero_box = _tui_tagged_box_lines(
+        width=hero_left_w, body_tagged=hero_body,
+        title="{accent.b}Current Week{/}",
+        pin=("{focused}focus{/}" if runtime.focus_index == 0 else "{dim}[1]{/}"),
+        border_style=("focused" if runtime.focus_index == 0 else "faint"),
+    )
+    trend_box = _tui_tagged_box_lines(
+        width=hero_right_w, body_tagged=trend_body,
+        title="{accent.b}$/1% Trend{/} {dim}· 8 weeks{/}",
+        pin=("{focused}focus{/}" if runtime.focus_index == 2 else "{dim}[3]{/}"),
+        border_style=("focused" if runtime.focus_index == 2 else "faint"),
+    )
+    m = max(len(hero_box), len(trend_box))
+
+    def _pad(lines: list[str], w: int, style: str) -> list[str]:
+        while len(lines) < m:
+            lines.insert(-1, f"{{{style}}}║{{/}}{' ' * (w - 2)}{{{style}}}║{{/}}")
+        return lines
+
+    hero_box = _pad(hero_box, hero_left_w, "focused" if runtime.focus_index == 0 else "faint")
+    trend_box = _pad(trend_box, hero_right_w, "focused" if runtime.focus_index == 2 else "faint")
+
+    # --- Forecast & Budget strip ----------------------------------------
+    if snap.forecast:
+        b100 = next((r for r in snap.forecast.budgets if r.target_percent == 100), None)
+        b90 = next((r for r in snap.forecast.budgets if r.target_percent == 90), None)
+        b100_s = (
+            f"${b100.dollars_per_day:.2f}/d"
+            if b100 and b100.dollars_per_day is not None else "—"
+        )
+        b90_s = (
+            f"${b90.dollars_per_day:.2f}/d"
+            if b90 and b90.dollars_per_day is not None else "—"
+        )
+        reset_str = (
+            format_display_dt(cw.week_end_at, runtime.display_tz, fmt="%b %d %H:%M", suffix=True)
+            if cw else "—"
+        )
+        fcstrip_body = [
+            "",
+            (f"  {{dim}}wk-avg{{/}} {{{_tui_bar_color(wa)}.b}}{wa}%{{/}}  "
+             f"{{dim}}24h{{/}} {{{_tui_bar_color(rc)}.b}}{rc}%{{/}}  "
+             f"{{faint}}│{{/}}  "
+             f"{{dim}}≤100%{{/}} {{bright.b}}{b100_s}{{/}}  "
+             f"{{dim}}≤90%{{/}} {{bright.b}}{b90_s}{{/}}  "
+             f"{{faint}}│{{/}}  "
+             f"{{dim}}reset{{/}} {{bright}}{reset_str}{{/}}"),
+            "",
+        ]
+    else:
+        fcstrip_body = ["", "  {dim}forecast unavailable{/}", ""]
+    fc_strip_box = _tui_tagged_box_lines(
+        width=width, body_tagged=fcstrip_body,
+        title="{accent.b}Forecast & Budget{/}",
+        pin=("{focused}focus{/}" if runtime.focus_index == 1 else "{dim}[2]{/}"),
+        border_style=("focused" if runtime.focus_index == 1 else "faint"),
+    )
+
+    # --- Sessions (full-width, always focused in B) --------------------
+    # Chrome around the sessions panel:
+    #   ribbon(1) + sub(1) + rule(1) + 1 blank
+    #   + hero_box lines
+    #   + 1 blank after hero
+    #   + fc_strip_box lines
+    #   + 1 blank after fc strip
+    #   + 2 lines for the sessions box borders (top + bottom)
+    #   + 5 sessions-panel chrome lines (leading blank + header + ruler
+    #     + trailing blank + "↑↓ scroll · N below")
+    #   + footer_lines(2)
+    if runtime.input_mode is not None:
+        match_count = None
+        if runtime.input_mode == "filter":
+            match_count = sum(
+                1 for s in snap.sessions
+                if (runtime.input_buffer.lower() in s.project_label.lower()
+                    or runtime.input_buffer.lower() in s.model_primary.lower())
+            ) if runtime.input_buffer else None
+        elif runtime.input_mode == "search":
+            if runtime.input_buffer:
+                needle = runtime.input_buffer.lower()
+                # See _tui_render_variant_a: count must match the navigable
+                # post-filter set used by _tui_panel_sessions.
+                pool = _tui_apply_session_filter(snap.sessions, runtime.filter_term)
+                count = 0
+                for s in pool:
+                    hay = (
+                        s.project_label.lower() + "|"
+                        + s.model_primary.lower() + "|"
+                        + _tui_format_started(s.started_at, snap.generated_at, runtime.display_tz).lower()
+                    )
+                    if needle in hay:
+                        count += 1
+                match_count = count
+            else:
+                match_count = None
+        footer_lines = _tui_render_input_prompt(runtime, width, match_count=match_count)
+    else:
+        footer_lines = _tui_footer_keys(width)
+    sessions_chrome = 5       # "" + header + ruler + "" + "↑↓ scroll · N below"
+    box_borders = 2           # sessions box top + bottom
+    non_session_rows = (
+        4                      # ribbon + sub + rule + blank
+        + len(hero_box)
+        + 1                    # blank after hero
+        + len(fc_strip_box)
+        + 1                    # blank after fc strip
+        + box_borders
+        + sessions_chrome
+        + len(footer_lines)
+    )
+    rows_visible = max(3, height - non_session_rows)
+    sess_body = _tui_panel_sessions(
+        snap, runtime, width,
+        rows_visible=rows_visible,
+        show_project_col=(bucket == "wide"),
+    )
+    sess_box = _tui_tagged_box_lines(
+        width=width, body_tagged=sess_body,
+        title=_tui_sessions_title(runtime, narrow=(bucket == "narrow")),
+        pin=("{focused}focus{/}" if runtime.focus_index == 3 else "{dim}[4]{/}"),
+        border_style=("focused" if runtime.focus_index == 3 else "faint"),
+    )
+
+    # --- Assemble -------------------------------------------------------
+    rule_line = "{faint}" + ("═" * width) + "{/}"
+    warn_line = (
+        "{warn}⚠ narrow terminal — some columns hidden{/}"
+        if bucket == "narrow" else None
+    )
+
+    hero_text = _tui_lines_to_text(hero_box)
+    trend_text = _tui_lines_to_text(trend_box)
+    fc_strip_text = _tui_lines_to_text(fc_strip_box)
+    sess_text = _tui_lines_to_text(sess_box)
+    footer_text = _tui_lines_to_text(footer_lines)
+
+    root = Layout()
+    regions: list[Layout] = [
+        Layout(name="ribbon", size=1),
+        Layout(name="sub", size=1),
+        Layout(name="rule", size=1),
+    ]
+    if warn_line is not None:
+        regions.append(Layout(name="warn", size=1))
+
+    if overlay_panel is not None:
+        # Fallback A: collapse the body bands into one ``body`` region
+        # holding a centered overlay Panel. Ribbon/sub/rule/footer remain.
+        body_rows = (
+            1                 # blank after rule
+            + len(hero_box)
+            + 1               # blank after hero
+            + len(fc_strip_box)
+            + 1               # blank after fc strip
+            + len(sess_box)
+        )
+        regions.append(Layout(name="body", size=body_rows))
+        regions.append(Layout(name="footer", size=len(footer_lines)))
+        root.split_column(*regions)
+        root["ribbon"].update(_tui_lines_to_text([ribbon]))
+        root["sub"].update(_tui_lines_to_text([sub]))
+        root["rule"].update(_tui_lines_to_text([rule_line]))
+        if warn_line is not None:
+            root["warn"].update(_tui_lines_to_text([warn_line]))
+        root["body"].update(Align.center(overlay_panel, vertical="middle"))
+        root["footer"].update(footer_text)
+        root._tui_natural_height = sum(r.size or 0 for r in regions)
+        return root
+
+    regions.extend([
+        Layout(name="blank1", size=1),
+        Layout(name="hero_row", size=len(hero_box)),
+        Layout(name="blank2", size=1),
+        Layout(name="fc_strip", size=len(fc_strip_box)),
+        Layout(name="blank3", size=1),
+        Layout(name="sessions", size=len(sess_box)),
+        Layout(name="footer", size=len(footer_lines)),
+    ])
+    root.split_column(*regions)
+
+    root["ribbon"].update(_tui_lines_to_text([ribbon]))
+    root["sub"].update(_tui_lines_to_text([sub]))
+    root["rule"].update(_tui_lines_to_text([rule_line]))
+    if warn_line is not None:
+        root["warn"].update(_tui_lines_to_text([warn_line]))
+
+    root["blank1"].update(_tui_lines_to_text([""]))
+    root["hero_row"].split_row(
+        Layout(name="hero", size=hero_left_w),
+        Layout(name="trend", size=hero_right_w),
+    )
+    root["hero_row"]["hero"].update(hero_text)
+    root["hero_row"]["trend"].update(trend_text)
+    root["blank2"].update(_tui_lines_to_text([""]))
+    root["fc_strip"].update(fc_strip_text)
+    root["blank3"].update(_tui_lines_to_text([""]))
+    root["sessions"].update(sess_text)
+    root["footer"].update(footer_text)
+    root._tui_natural_height = sum(r.size or 0 for r in regions)
+    return root
+
+
+_TUI_HELP_LINES = [
+    "",
+    "  {accent.b}Dashboard{/}",
+    "",
+    "    {bright}q{/} {dim}·{/} {fg}quit{/}",
+    "    {bright}r{/} {dim}·{/} {fg}force refresh{/}",
+    "    {bright}v{/} {dim}·{/} {fg}toggle variant (conventional/expressive){/}",
+    "    {bright}Tab{/} {dim}·{/} {fg}cycle focus across panels{/}",
+    "    {bright}↑↓ / j k{/} {dim}·{/} {fg}scroll sessions or modal{/}",
+    "    {bright}PgUp/PgDn{/} {dim}·{/} {fg}page scroll{/}",
+    "",
+    "  {accent.b}Sessions panel{/}",
+    "    {bright}s{/} {dim}·{/} {fg}cycle sort key{/}",
+    "    {bright}f{/} {dim}·{/} {fg}filter (project|model substring){/}",
+    "    {bright}/{/} {dim}·{/} {fg}search (highlight + jump){/}",
+    "    {bright}n / N{/} {dim}·{/} {fg}next/prev search match{/}",
+    "",
+    "  {accent.b}Detail modals{/}",
+    "    {bright}Enter{/} {dim}·{/} {fg}open detail of focused panel{/}",
+    "    {bright}1 2 3 4{/} {dim}·{/} {fg}open Current/Forecast/Trend/Sessions detail{/}",
+    "    {bright}Esc{/} {dim}·{/} {fg}close modal · cancel input · close help{/}",
+    "",
+    "  {bright}?{/} {dim}·{/} {fg}toggle help{/}",
+    "",
+]
+
+
+def _tui_render_help(width: int, height: int) -> Panel:
+    """Return a ``rich.panel.Panel`` for the help overlay.
+
+    The Panel lists the keybindings; the caller is responsible for
+    centering it via ``rich.align.Align.center`` and composing it over
+    the variant Layout (see the body-region-swap overlay in
+    ``_tui_render_variant_a`` / ``_tui_render_variant_b``).
+
+    ``height`` is accepted for API symmetry but not currently used —
+    the Panel auto-sizes to its content.
+    """
+    from rich import box as _rich_box
+    from rich.panel import Panel as _Panel
+
+    # Build body Text from _TUI_HELP_LINES verbatim (tags unchanged).
+    body = _tui_lines_to_text(_TUI_HELP_LINES)
+    panel_w = min(max(width - 4, 20), 60)
+    return _Panel(
+        body,
+        box=_rich_box.DOUBLE,
+        title=_tui_colortag("{accent.b}Help{/}"),
+        subtitle=_tui_colortag("{dim}? to close{/}"),
+        border_style="accent",
+        width=panel_w,
+    )
+
+
+def _tui_modal_max_width(width: int) -> int:
+    """Per-bucket modal width (spec §5.1)."""
+    bucket = _tui_width_bucket(width)
+    if bucket == "wide":
+        return min(width - 4, 90)
+    if bucket == "compact":
+        return min(width - 4, 70)
+    # narrow
+    return max(60, width - 2)
+
+
+def _tui_render_modal(
+    snap: DataSnapshot,
+    runtime: RuntimeState,
+    width: int,
+    height: int,
+) -> Panel:
+    """Render the active detail modal as a centered Panel.
+
+    Dispatches on runtime.modal_kind to a per-kind content builder.
+    Per-kind builders return list[str] (color-tagged lines); this
+    function slices them by runtime.modal_scroll, wraps in a Panel
+    with the shared chrome, and returns it for body-region-swap
+    composition.
+
+    Spec §4.1 (chrome) + §4.6 (per-kind content).
+    """
+    from rich import box as _rich_box
+    from rich.panel import Panel as _Panel
+
+    kind = runtime.modal_kind or "current_week"
+    if kind == "current_week":
+        title, content_lines = _tui_modal_current_week(snap, runtime, width)
+    elif kind == "forecast":
+        title, content_lines = _tui_modal_forecast(snap, runtime, width)
+    elif kind == "trend":
+        title, content_lines = _tui_modal_trend(snap, runtime, width)
+    elif kind == "session":
+        title, content_lines = _tui_modal_session(snap, runtime, width)
+    else:
+        title = "Modal"
+        content_lines = ["", "  {dim}placeholder{/}", ""]
+
+    panel_w = _tui_modal_max_width(width)
+    panel_h = min(height - 4, 30)
+    viewport = max(5, panel_h - 4)  # subtract title + subtitle + padding
+
+    total = len(content_lines)
+    scroll = max(0, min(runtime.modal_scroll, max(0, total - viewport)))
+    runtime.modal_scroll = scroll  # write back the clamp
+    visible = content_lines[scroll : scroll + viewport]
+    # Pad to viewport so the panel height is stable across scrolls.
+    while len(visible) < viewport:
+        visible.append("")
+
+    body = _tui_lines_to_text(visible)
+    subtitle = "{dim}Esc back{/}"
+    if total > viewport:
+        subtitle = f"{{dim}}Esc back · {scroll + 1}-{scroll + viewport}/{total} ↓{{/}}"
+
+    return _Panel(
+        body,
+        box=_rich_box.DOUBLE,
+        title=_tui_colortag(title),
+        subtitle=_tui_colortag(subtitle),
+        border_style="accent",
+        width=panel_w,
+    )
+
+
+# ---- per-kind modal content builders (spec §4.6) ----
+def _tui_modal_current_week(snap, runtime, width):
+    """Per-percent milestones for the current week (spec §4.6.1)."""
+    cw = snap.current_week
+    milestones = snap.percent_milestones
+    if cw is None:
+        return ("{accent.b}Current Week · per-percent{/}",
+                ["", "  {dim}No current week — run record-usage{/}", ""])
+    if not milestones:
+        return ("{accent.b}Current Week · per-percent{/}",
+                ["", "  {dim}No milestones yet — keep recording usage.{/}", ""])
+    avg_dpp = (cw.dollars_per_percent or 0.0)
+    cumul = cw.spent_usd
+    header = [
+        "",
+        f"  {{dim}}Week{{/}} {{b}}{format_display_dt(cw.week_start_at, runtime.display_tz, fmt='%b %d', suffix=False)} – {format_display_dt(cw.week_end_at, runtime.display_tz, fmt='%b %d', suffix=False)}{{/}}   "
+        f"{{dim}}milestones reached{{/}} {{warn.b}}{len(milestones)}{{/}}",
+        f"  {{dim}}avg $/1%{{/}} {{b}}${avg_dpp:.2f}{{/}}     {{dim}}cumulative{{/}} {{b}}${cumul:.2f}{{/}}",
+        "",
+    ]
+    bucket = _tui_width_bucket(width)
+    show_5h = bucket != "narrow"
+    if show_5h:
+        header.append("   {dim.b}  %  Crossed at             Cumul    Marginal   5-hr{/}")
+        header.append("   {faint}─── ────────────────────── ──────── ────────── ──────{/}")
+    else:
+        header.append("   {dim.b}  %  Crossed at             Cumul    Marginal{/}")
+        header.append("   {faint}─── ────────────────────── ──────── ──────────{/}")
+    rows = []
+    for ms in milestones:
+        ts_str = format_display_dt(
+            ms.crossed_at, runtime.display_tz,
+            fmt="%b %d %H:%M:%S", suffix=True,
+        )
+        cumul_str = f"${ms.cumulative_cost_usd:.2f}".ljust(8)
+        marg_str = (f"${ms.marginal_cost_usd:.2f}" if ms.marginal_cost_usd is not None else "—").ljust(10)
+        line = f"   {{b}}{ms.percent:>3}{{/}} {{bright}}{ts_str:<22}{{/}} {{b}}{cumul_str}{{/}} {{b}}{marg_str}{{/}}"
+        if show_5h:
+            five_str = (f"{int(ms.five_hour_pct_at_crossing)}%"
+                        if ms.five_hour_pct_at_crossing is not None else "—")
+            line += f" {{dim}}{five_str:<5}{{/}}"
+        rows.append(line)
+    if runtime.modal_snap_pending:
+        if len(rows) > 10:
+            runtime.modal_scroll = len(rows) + len(header) - 10
+        runtime.modal_snap_pending = False
+    return ("{accent.b}Current Week · per-percent{/}", header + rows)
+
+def _tui_modal_forecast(snap, runtime, width):
+    """Forecast --explain content (spec §4.6.2)."""
+    fc = snap.forecast
+    if fc is None or getattr(fc, "inputs", None) is None:
+        return ("{accent.b}Forecast · explain{/}",
+                ["", "  {dim}Forecast unavailable — current week is empty.{/}", ""])
+    inp = fc.inputs
+    verdict = _tui_verdict_of(fc)
+    vcls = _TUI_VERDICT_CLS[verdict]
+    # Hero label band right-pads to 15 chars so value columns align:
+    # "Now" + 12, "Week elapsed" + 3, "Used now" + 7, "Used 24h ago" + 3.
+    lines = [
+        "",
+        f"  {{dim}}Now{{/}}            {{b}}{format_display_dt(inp.now_utc, runtime.display_tz, fmt='%Y-%m-%d %H:%M', suffix=True)}{{/}}",
+        f"  {{dim}}Week elapsed{{/}}   {{b}}{inp.elapsed_hours:.1f}h / 168h{{/}}  "
+        f"{{dim}}({(inp.elapsed_hours / 168 * 100):.1f}%){{/}}",
+        f"  {{dim}}Used now{{/}}       {{warn.b}}{inp.p_now:.1f}%{{/}}",
+    ]
+    if inp.p_24h_ago is not None:
+        lines.append(f"  {{dim}}Used 24h ago{{/}}   {{dim}}{inp.p_24h_ago:.1f}%{{/}}")
+    else:
+        lines.append("  {dim}Used 24h ago{/}   {dim}—  (insufficient history){/}")
+    lines.append("")
+    lines.append("  {dim.b}Two rate paths{/}")
+    lines.append(f"  {{dim}}  r_avg     {inp.p_now:.1f} / {inp.elapsed_hours:.1f}      = {{/}}{{b}}{fc.r_avg:.4f} %/h{{/}}")
+    if fc.r_recent is not None and inp.p_24h_ago is not None:
+        lines.append(f"  {{dim}}  r_recent  ({inp.p_now:.1f}-{inp.p_24h_ago:.1f}) / {inp.t_24h_actual_hours:.1f}  = {{/}}{{b}}{fc.r_recent:.4f} %/h{{/}}")
+    else:
+        lines.append("  {dim}  r_recent  unavailable — no 24h-prior sample{/}")
+    lines.append("")
+    lines.append(f"  {{dim.b}}Project to week end ({inp.remaining_hours:.1f}h remaining){{/}}")
+    if fc.r_recent is not None:
+        wa = inp.p_now + fc.r_avg * inp.remaining_hours
+        rc = inp.p_now + fc.r_recent * inp.remaining_hours
+        lines.append(f"  {{dim}}  by week-avg    = {{/}}{{warn}}{wa:.1f}%{{/}}")
+        lines.append(f"  {{dim}}  by recent-24h  = {{/}}{{ok}}{rc:.1f}%{{/}}")
+        lines.append(f"  {{dim}}  high           = {{/}}{{{vcls}.b}}{fc.final_percent_high:.1f}%{{/}}     {{dim}}verdict:{{/}} {{{vcls}.b}}{verdict}{{/}}")
+    else:
+        lines.append(f"  {{dim}}  projection     = {{/}}{{{vcls}.b}}{fc.final_percent_high:.1f}%{{/}}     {{dim}}verdict:{{/}} {{{vcls}.b}}{verdict}{{/}}")
+    lines.append("")
+    lines.append(f"  {{dim.b}}Daily $ budgets ({inp.remaining_days:.3f} days remaining){{/}}")
+    for b in fc.budgets:
+        if b.dollars_per_day is not None:
+            lines.append(f"  {{dim}}  ≤{b.target_percent}%   {{/}}{{b}}${b.dollars_per_day:.2f}/day{{/}}")
+        else:
+            lines.append(f"  {{dim}}  ≤{b.target_percent}%   {{/}}{{dim}}—  (already past){{/}}")
+    lines.append("")
+    confidence = inp.confidence
+    lines.append(f"  {{dim}}confidence: {confidence} · based on 7-day rate{{/}}")
+    return ("{accent.b}Forecast · explain{/}", lines)
+
+def _tui_modal_trend(snap, runtime, width):
+    """Weekly history for the Trend modal (spec §4.6.3)."""
+    history = snap.weekly_history
+    if not history:
+        return ("{accent.b}Trend · weekly history{/}",
+                ["", "  {dim}No weekly history available yet.{/}", ""])
+    bucket = _tui_width_bucket(width)
+    show_age = bucket != "narrow"
+    # Header
+    valid = [h for h in history if h.dollars_per_percent is not None]
+    avg_dpp = sum(h.dollars_per_percent for h in valid) / len(valid) if valid else 0.0
+    if len(valid) >= 2:
+        first_dpp = valid[0].dollars_per_percent
+        last_dpp = valid[-1].dollars_per_percent
+        trend_pct = ((last_dpp - first_dpp) / first_dpp * 100) if first_dpp else 0.0
+    else:
+        trend_pct = 0.0
+    trend_sign = "+" if trend_pct >= 0 else ""
+    trend_cls = "warn" if abs(trend_pct) >= 5 else "ok"
+    lines = [
+        "",
+        f"  {{dim}}Last{{/}} {{b}}{len(history)} weeks{{/}}     "
+        f"{{dim}}avg $/1%{{/}} {{b}}${avg_dpp:.2f}{{/}}     "
+        f"{{dim}}trend{{/}} {{{trend_cls}}}{trend_sign}{trend_pct:.0f}%{{/}}",
+        "",
+    ]
+    if show_age:
+        lines.append("   {dim.b}  Week starting    Used%    Cost     $/1%      Δ{/}")
+        lines.append("   {faint}─────────────────  ──────  ────────  ──────  ──────{/}")
+    else:
+        lines.append("   {dim.b}  Week     Used%    Cost     $/1%{/}")
+        lines.append("   {faint}─────────  ──────  ────────  ──────{/}")
+    n = len(history)
+    for i, h in enumerate(history):
+        ago = n - 1 - i
+        marker = "▶" if h.is_current else " "
+        if h.used_pct is None:
+            used_cell = "—"
+            used_cls = "dim"
+        else:
+            used_cell = f"{h.used_pct:.1f}%"
+            used_cls = "ok" if h.used_pct < 70 else ("warn" if h.used_pct < 90 else "bad")
+        dpp_cell = f"${h.dollars_per_percent:.2f}" if h.dollars_per_percent is not None else "—"
+        delta_cell = ""
+        if h.delta_dpp is not None:
+            sign = "+" if h.delta_dpp >= 0 else ""
+            cls = "ok" if h.delta_dpp >= 0 else "dim"
+            delta_cell = f"{{{cls}}}{sign}{h.delta_dpp:.2f}{{/}}"
+        else:
+            delta_cell = "{dim}  —{/}"
+        cost_cell = (f"${(h.dollars_per_percent or 0) * (h.used_pct or 0):.2f}"
+                     if h.used_pct is not None and h.dollars_per_percent else "—")
+        if show_age:
+            label = f"{h.week_label} ({ago:>2}w ago)" if ago > 0 else f"{h.week_label} (now)"
+            lines.append(
+                f"   {{focused.b}}{marker}{{/}} {{b}}{label:<17}{{/}}  "
+                f"{{{used_cls}}}{used_cell:>5}{{/}}  {{b}}{cost_cell:>7}{{/}}  "
+                f"{{b}}{dpp_cell:>5}{{/}}   {delta_cell}"
+            )
+        else:
+            lines.append(
+                f"   {{focused.b}}{marker}{{/}} {{b}}{h.week_label:<7}{{/}}  "
+                f"{{{used_cls}}}{used_cell:>5}{{/}}  {{b}}{cost_cell:>7}{{/}}  {{b}}{dpp_cell:>5}{{/}}"
+            )
+    lines.append("")
+    # Default scroll: bottom (current week is most relevant).
+    if runtime.modal_snap_pending:
+        if len(lines) > 12:
+            runtime.modal_scroll = len(lines) - 12
+        runtime.modal_snap_pending = False
+    return ("{accent.b}Trend · weekly history{/}", lines)
+
+def _tui_modal_session(snap, runtime, width):
+    """Session detail modal (spec §4.6.4).
+
+    Looks up the topmost-visible session by index and queries
+    _tui_build_session_detail on demand. Fixture goldens may inject
+    a deterministic detail via runtime.session_detail_override.
+    """
+    sessions = _tui_sort_sessions(snap.sessions, runtime.sort_key)
+    if runtime.filter_term:
+        af_lower = runtime.filter_term.lower()
+        sessions = [
+            s for s in sessions
+            if af_lower in s.project_label.lower()
+            or af_lower in s.model_primary.lower()
+        ]
+    if not sessions:
+        return ("{accent.b}Session · detail{/}",
+                ["", "  {dim}No session selected.{/}", ""])
+    idx = max(0, min(runtime.session_scroll, len(sessions) - 1))
+    sel = sessions[idx]
+    # Fixture-injection hook (dev-only, spec §5.5)
+    detail = getattr(runtime, "session_detail_override", None)
+    if detail is None:
+        cache = runtime.session_detail_cache
+        if (cache is not None
+                and cache[0] == sel.session_id
+                and cache[1] == snap.generated_at):
+            detail = cache[2]
+        else:
+            detail = _tui_build_session_detail(sel.session_id, now_utc=snap.generated_at)
+            runtime.session_detail_cache = (sel.session_id, snap.generated_at, detail)
+    if detail is None:
+        return ("{accent.b}Session · detail{/}",
+                ["", "  {warn}Session no longer available · Esc to return{/}", ""])
+    bucket = _tui_width_bucket(width)
+    show_cwd = bucket != "narrow"
+    show_full_id = bucket != "narrow"
+    sid_display = detail.session_id if show_full_id else detail.session_id[:8]
+
+    title = f"{{accent.b}}Session · {format_display_dt(detail.started_at, runtime.display_tz, fmt='%H:%M:%S', suffix=True)} ({_tui_escape_tags(detail.project_label)}){{/}}"
+    lines = [
+        "",
+        f"  {{dim}}Session ID{{/}}     {{b}}{sid_display}{{/}}",
+        f"  {{dim}}Started{{/}}        {{b}}{format_display_dt(detail.started_at, runtime.display_tz, fmt='%Y-%m-%d %H:%M:%S', suffix=True)}{{/}}",
+        f"  {{dim}}Last activity{{/}}  {{b}}{format_display_dt(detail.last_activity_at, runtime.display_tz, fmt='%Y-%m-%d %H:%M:%S', suffix=True)}{{/}}",
+        f"  {{dim}}Duration{{/}}       {{b}}{_tui_format_dur(detail.duration_minutes)}{{/}}",
+        f"  {{dim}}Project{{/}}        {{b}}{_tui_escape_tags(detail.project_label)}{{/}}",
+    ]
+    if show_cwd:
+        cwd_max = max(20, _tui_modal_max_width(width) - 18)
+        cwd_shown = detail.project_path
+        if len(cwd_shown) > cwd_max:
+            cwd_shown = "…" + cwd_shown[-(cwd_max - 1):]
+        lines.append(f"  {{dim}}  cwd{{/}}          {{dim}}{_tui_escape_tags(cwd_shown)}{{/}}")
+    src_count = len(detail.source_paths)
+    src_note = "1 (no resumes across files)" if src_count == 1 else f"{src_count} (resumed across files)"
+    lines.append(f"  {{dim}}Source files{{/}}   {{b}}{src_note}{{/}}")
+    lines.append("")
+    lines.append("  {dim.b}Models{/}")
+    for model_name, role in detail.models:
+        padded = f"{model_name:<16}"
+        lines.append(f"  {{dim}}  {{/}}{{b}}{_tui_escape_tags(padded)}{{/}}{{dim}}{role}{{/}}")
+    lines.append("")
+    lines.append("  {dim.b}Tokens{/}")
+    lines.append(f"  {{dim}}  Input        {{/}} {{b}}{detail.input_tokens:>10,}{{/}}")
+    lines.append(f"  {{dim}}  Cache create {{/}} {{b}}{detail.cache_creation_tokens:>10,}{{/}}")
+    cache_pct_str = (f"   {{ok}}{int(detail.cache_hit_pct)}% cache hit{{/}}"
+                     if detail.cache_hit_pct is not None else "")
+    lines.append(f"  {{dim}}  Cache read   {{/}} {{b}}{detail.cache_read_tokens:>10,}{{/}}{cache_pct_str}")
+    lines.append(f"  {{dim}}  Output       {{/}} {{b}}{detail.output_tokens:>10,}{{/}}")
+    lines.append("")
+    lines.append("  {dim.b}Cost{/}")
+    for model_name, cost in detail.cost_per_model:
+        padded = f"{model_name:<13}"
+        lines.append(f"  {{dim}}  {_tui_escape_tags(padded)}{{/}} {{b}}${cost:.2f}{{/}}")
+    lines.append("  {faint}  ─────────────────────{/}")
+    lines.append(f"  {{dim}}  Total         {{/}} {{b}}${detail.cost_total_usd:.2f}{{/}}")
+    return (title, lines)
+
+
+def _tui_render_toast(msg: str, width: int):
+    """Render a one-line deferred-feature toast as a rich.text.Text.
+
+    The toast surfaces the message in a warn-badge style. `width` is
+    accepted for symmetry with the other renderers but is not used for
+    padding — the toast sits inline wherever the caller places it.
+    """
+    content = f"  {msg}  "
+    padded = f" {{badge.warn}}{content}{{/}} "
+    return _tui_colortag(padded)
+
+
+def _tui_sync_interval_type(s: str) -> float:
+    """argparse type validator for --sync-interval: float >= 1.0."""
+    try:
+        v = float(s)
+    except ValueError:
+        raise argparse.ArgumentTypeError(
+            f"--sync-interval must be a number (got {s!r})"
+        )
+    if v < 1.0:
+        raise argparse.ArgumentTypeError(
+            f"--sync-interval must be >= 1.0 seconds (got {v})"
+        )
+    return v
+
+
+def _tui_refresh_interval_type(s: str) -> float:
+    """argparse type validator for --refresh: float > 0.0.
+
+    Non-positive values would make the keyboard-poll select() return
+    immediately every iteration, busy-spinning the redraw loop.
+    """
+    try:
+        v = float(s)
+    except ValueError:
+        raise argparse.ArgumentTypeError(
+            f"--refresh must be a number (got {s!r})"
+        )
+    if v <= 0.0:
+        raise argparse.ArgumentTypeError(
+            f"--refresh must be > 0 seconds (got {v})"
+        )
+    return v
+
+
+def _make_run_sync_now_locked(*, ref, hub, pinned_now, display_tz_pref_override):
+    """Return a closure that does the snapshot-rebuild + SSE-publish work.
+
+    Caller MUST hold sync_lock around the call. The naming convention
+    (``_locked`` suffix) is the contract; threading.Lock has no
+    "is_held_by_current_thread" check, so we don't introspect.
+
+    Splitting the locked body out of the public wrapper lets ``/api/sync``
+    callers that already hold ``sync_lock`` (e.g. so they can refresh OAuth
+    + rebuild snapshot atomically without releasing between steps) reuse
+    this body without recursive-acquire / self-deadlock.
+    """
+    def _locked(skip_sync: bool) -> None:
+        try:
+            # Resolve _tui_build_snapshot via cctally's namespace so the
+            # eager re-export AND ``monkeypatch.setitem(ns, "_tui_build_snapshot", spy)``
+            # in tests/test_dashboard_api_sync_refresh.py propagate into
+            # this closure body (the bare-name lookup would resolve in
+            # this sibling's __dict__ and miss the cctally-side patch).
+            snap = sys.modules["cctally"]._tui_build_snapshot(
+                now_utc=pinned_now, skip_sync=skip_sync,
+                display_tz_pref_override=display_tz_pref_override,
+            )
+            if skip_sync:
+                # Mirror the startup override: suppress the monotonic sync
+                # stamp so the envelope keeps emitting sync_age_s=None and
+                # the client keeps rendering "sync paused" after the user
+                # hits r / clicks the sync chip.
+                snap = dataclasses.replace(snap, last_sync_at=None)
+            ref.set(snap)
+            hub.publish(snap)
+        except Exception as exc:
+            prev = ref.get()
+            crashed = dataclasses.replace(
+                prev,
+                last_sync_error=f"sync crashed: {exc}",
+                generated_at=dt.datetime.now(dt.timezone.utc),
+            )
+            ref.set(crashed)
+            hub.publish(crashed)
+    return _locked
+
+
+def _make_run_sync_now(*, sync_lock, ref, hub, pinned_now,
+                       display_tz_pref_override):
+    """Return a closure that acquires sync_lock then runs the locked variant.
+
+    Used by the periodic background thread and by anything else that needs
+    full lifecycle (acquire-do-release) in one call. ``/api/sync`` paths
+    that compose multiple lock-protected steps should use the locked variant
+    directly instead of nesting ``with sync_lock:`` (re-entrant acquire on
+    a non-recursive ``threading.Lock`` self-deadlocks).
+    """
+    locked = _make_run_sync_now_locked(
+        ref=ref, hub=hub, pinned_now=pinned_now,
+        display_tz_pref_override=display_tz_pref_override,
+    )
+    def _public(skip_sync: bool) -> None:
+        with sync_lock:
+            locked(skip_sync)
+    return _public
+
+
+
+def cmd_tui(args: argparse.Namespace) -> int:
+    """Launch the live TUI dashboard. See docs/commands/tui.md.
+
+    Live-path state machine:
+      1. Resolve `now_utc` (honor --as-of).
+      2. Build RuntimeState from args (honors NO_COLOR + --no-color).
+      3. If --render-once: defer to `_tui_render_once` and return.
+      4. Build theme, construct Console(theme, no_color=...).
+      5. Refuse if terminal width < 80 columns.
+      6. Build initial snapshot (skip_sync=True, non-blocking).
+      7. Wrap in _SnapshotRef.
+      8. Start _TuiSyncThread unless --no-sync.
+      9. rich.live.Live with alternate screen, auto_refresh=False.
+     10. TuiKeyReader (raw mode, cbreak).
+     11. SIGINT → should_exit flag (no raising into Live).
+         SIGWINCH → no-op (next tick picks up console.size).
+     12. On every tick: read key → mutate runtime → live.update().
+         On no key: natural tick redraw.
+     13. Finally: stop sync thread.
+    """
+    try:
+        import rich  # noqa: F401
+    except ImportError:
+        print(TUI_RICH_MISSING_MSG, file=sys.stderr)
+        return 1
+
+    # --- 1. Resolve now ----------------------------------------------------
+    now_utc = _resolve_forecast_now(getattr(args, "as_of", None))
+
+    # --- 2a. Resolve display tz via the unified --tz / config.display.tz.
+    # RuntimeState.tz keeps the legacy token shape for any string-keyed
+    # call sites; F4 moved _tui_format_started to consume display_tz
+    # (ZoneInfo | None) directly so non-"local" values now localize
+    # correctly instead of falling back to UTC. Normalize args.tz back to
+    # a token shape: None -> "local"; Etc/UTC -> "utc"; explicit IANA ->
+    # the verbatim IANA name.
+    config = load_config()
+    # Capture the raw `--tz` flag BEFORE resolution rewrites args.tz, so
+    # `_tui_build_snapshot` can apply the same persisted-config override
+    # that `cmd_dashboard` uses (parallel to lines 24927-24936). Without
+    # this, panels that precompute labels at snapshot-build time (trend,
+    # weekly-history) render the persisted `config.display.tz` instead of
+    # honoring the explicit per-call `--tz` override.
+    raw_tz_flag = getattr(args, "tz", None)
+    if raw_tz_flag is not None and str(raw_tz_flag).strip() != "":
+        try:
+            display_tz_pref_override = normalize_display_tz_value(raw_tz_flag)
+        except ValueError:
+            display_tz_pref_override = None
+    else:
+        display_tz_pref_override = None
+    tz_obj = resolve_display_tz(args, config)
+    args._resolved_tz = tz_obj
+    if tz_obj is None:
+        args.tz = "local"
+    elif tz_obj.key == "Etc/UTC":
+        args.tz = "utc"
+    else:
+        args.tz = tz_obj.key
+    # Stash the override on `args` so `_tui_render_once` (the dev path)
+    # can pick it up uniformly without a separate kwarg.
+    args._display_tz_pref_override = display_tz_pref_override
+
+    # --- 2. Runtime state --------------------------------------------------
+    runtime = RuntimeState.initial(args)
+
+    # --- 3. Dev path: one-shot render -------------------------------------
+    if getattr(args, "render_once", False):
+        return _tui_render_once(args, runtime, now_utc=now_utc)
+
+    # --- 3b. Require an interactive terminal ------------------------------
+    # Live mode drives alt-screen via rich.live.Live(screen=True) and reads
+    # keys from stdin in cbreak mode. Without a TTY on both ends, there is
+    # no quit path short of SIGINT and Live's escape sequences are
+    # meaningless. Refuse fast so cron/CI invocations fail instead of
+    # wedging. `--render-once` (above) is the scriptable alternative.
+    if not sys.stdin.isatty() or not sys.stdout.isatty():
+        print(
+            "tui: requires an interactive terminal "
+            "(stdin and stdout must be TTYs). "
+            "For scripted use, try `report`, `forecast`, or "
+            "`tui --render-once --snapshot-module PATH`.",
+            file=sys.stderr,
+        )
+        return 2
+
+    # --- 4. Theme ----------------------------------------------------------
+    # Drift guard moved to module scope (see _TUI_THEME_KEYS) and to
+    # _tui_build_theme itself, so both axes fire at import time / theme
+    # construction rather than on subcommand invocation.
+    theme = _tui_build_theme()
+
+    # --- 5. Console + width refuse ----------------------------------------
+    from rich.console import Console
+    from rich.live import Live
+    console = Console(theme=theme, no_color=not runtime.color_enabled)
+    width = console.size.width
+    if _tui_width_bucket(width) == "refuse":
+        print(
+            f"tui: terminal too narrow, need >=80 cols (got {width})",
+            file=sys.stderr,
+        )
+        return 1
+
+    # --- 6. Initial snapshot ----------------------------------------------
+    try:
+        initial_snap = _tui_build_snapshot(
+            now_utc=now_utc, skip_sync=True,
+            display_tz_pref_override=display_tz_pref_override,
+        )
+    except Exception:
+        initial_snap = _tui_empty_snapshot(now_utc)
+
+    # --- 7. Shared ref -----------------------------------------------------
+    ref = _SnapshotRef(initial_snap)
+
+    # --- 11. Signal handlers ----------------------------------------------
+    # Install signal handlers BEFORE sync.start() so a SIGINT during
+    # thread startup is caught by our flag-setter rather than the default
+    # handler (which would unwind past the finally block that calls
+    # sync.stop(), leaking a daemon thread). See Task 26 review I3.
+    import signal
+    import time as _time
+    should_exit = {"flag": False}
+
+    def _on_sigint(_signum, _frame):
+        should_exit["flag"] = True
+
+    prev_sigint = signal.getsignal(signal.SIGINT)
+    signal.signal(signal.SIGINT, _on_sigint)
+    prev_sigwinch = None
+    if hasattr(signal, "SIGWINCH"):
+        # No-op — Live's next tick reads console.size fresh.
+        prev_sigwinch = signal.getsignal(signal.SIGWINCH)
+        signal.signal(signal.SIGWINCH, lambda *_: None)
+    # SIGCONT (resume from Ctrl-Z): same idea — just let next tick redraw.
+    prev_sigcont = None
+    if hasattr(signal, "SIGCONT"):
+        prev_sigcont = signal.getsignal(signal.SIGCONT)
+        signal.signal(signal.SIGCONT, lambda *_: None)
+
+    # --- 8. Sync thread ----------------------------------------------------
+    # Always start the rebuild thread — even under --no-sync we need
+    # periodic refreshes so countdowns, "synced Xs ago", and external DB
+    # writes keep the dashboard live. The thread's own skip_sync flag
+    # gates the JSONL ingest pass; the rebuild itself is always cheap
+    # (SQLite SELECTs only, no JSONL scan when skip_sync=True).
+    # Only pin the sync thread's clock when a clock override was actually
+    # supplied — either via --as-of (CLI) or CCTALLY_AS_OF (env). Without
+    # one of those, `now_utc` is just "now captured once at startup";
+    # feeding that to the sync thread would freeze every subsequent
+    # rebuild on that instant. Mirroring the same check that
+    # _resolve_forecast_now() performed above keeps the hidden test hook
+    # consistent across the first frame and every subsequent tick.
+    pinned_now = now_utc if (
+        getattr(args, "as_of", None) or os.environ.get("CCTALLY_AS_OF")
+    ) else None
+    sync = _TuiSyncThread(
+        ref, float(args.sync_interval),
+        skip_sync=bool(getattr(args, "no_sync", False)),
+        now_utc=pinned_now,
+        display_tz_pref_override=display_tz_pref_override,
+    )
+    sync.start()
+
+    # --- 10. Render closure -----------------------------------------------
+    from rich.console import Group
+    def render():
+        snap = ref.get()
+        w = console.size.width
+        h = console.size.height
+        bucket = _tui_width_bucket(w)
+        # Build the help Panel once per render so both variants can receive
+        # it via the body-region-swap overlay path (Fallback A). When
+        # show_help is False, help_panel stays None and the variants render
+        # the normal 2x2 / hero layout.
+        help_panel = _tui_render_help(w, h) if runtime.show_help else None
+        modal_panel = (_tui_render_modal(snap, runtime, w, h)
+                       if runtime.modal_kind else None)
+        # v2: modal wins when both would show. Spec §4.7.
+        overlay = modal_panel or help_panel
+        if runtime.variant == "expressive":
+            frame = _tui_render_variant_b(
+                snap, runtime, w, h, bucket, overlay_panel=overlay,
+            )
+        else:
+            frame = _tui_render_variant_a(
+                snap, runtime, w, h, bucket, overlay_panel=overlay,
+            )
+        # Toast handling: expire when clock passes expiry; else stack below
+        # the frame via ``rich.console.Group`` (the rich-native stacking
+        # primitive required by constraint #4).
+        if runtime.toast is not None:
+            msg, expiry = runtime.toast
+            if _time.monotonic() < expiry:
+                toast_frame = _tui_render_toast(msg, w)
+                return Group(frame, toast_frame)
+            runtime.toast = None
+        return frame
+
+    # --- 12. Main loop ----------------------------------------------------
+    reader = TuiKeyReader()
+    try:
+        with reader, Live(
+            render(), console=console, screen=True,
+            auto_refresh=False, transient=False,
+        ) as live:
+            while not should_exit["flag"]:
+                key = reader.read(timeout=float(args.refresh))
+                if key is not None:
+                    redraw, quit_ = _tui_handle_key(key, runtime, ref)
+                    if quit_:
+                        break
+                    if redraw:
+                        live.update(render(), refresh=True)
+                        continue
+                # No key (or key with no redraw) — natural tick redraw.
+                live.update(render(), refresh=True)
+    finally:
+        # --- 13. Teardown -------------------------------------------------
+        if sync is not None:
+            sync.stop()
+        # Restore previous signal handlers.
+        try:
+            signal.signal(signal.SIGINT, prev_sigint)
+        except Exception:
+            pass
+        if hasattr(signal, "SIGWINCH") and prev_sigwinch is not None:
+            try:
+                signal.signal(signal.SIGWINCH, prev_sigwinch)
+            except Exception:
+                pass
+        if hasattr(signal, "SIGCONT") and prev_sigcont is not None:
+            try:
+                signal.signal(signal.SIGCONT, prev_sigcont)
+            except Exception:
+                pass
+    return 0
+
+
+def _tui_render_once(
+    args: argparse.Namespace,
+    runtime: "RuntimeState",
+    *,
+    now_utc: dt.datetime | None = None,
+) -> int:
+    """Dev-only: render one frame and emit plain text to stdout.
+
+    Used by fixture goldens (later Tasks 28-29). Honors:
+      --snapshot-module (load SNAPSHOT from a Python module for deterministic data)
+      --force-size WxH  (default 120x36 if unset / malformed)
+
+    Does NOT check the width-refuse bucket — --render-once is a dev path
+    expected to work at any size so authors can capture narrow-width goldens.
+    Returns 0 on success, 2 on malformed --force-size.
+    """
+    now_utc = now_utc or _resolve_forecast_now(getattr(args, "as_of", None))
+
+    # --- Parse --force-size -----------------------------------------------
+    force_size = getattr(args, "force_size", None)
+    w, h = 120, 36
+    if force_size:
+        parts = force_size.lower().split("x", 1)
+        if len(parts) != 2:
+            print(
+                f"tui: --force-size must be WxH (got {force_size!r})",
+                file=sys.stderr,
+            )
+            return 2
+        try:
+            w = int(parts[0])
+            h = int(parts[1])
+        except ValueError:
+            print(
+                f"tui: --force-size W/H must be integers (got {force_size!r})",
+                file=sys.stderr,
+            )
+            return 2
+        if w <= 0 or h <= 0:
+            print(
+                f"tui: --force-size W/H must be positive (got {force_size!r})",
+                file=sys.stderr,
+            )
+            return 2
+
+    # --- Load snapshot ----------------------------------------------------
+    snapshot_module = getattr(args, "snapshot_module", None)
+    snap: DataSnapshot
+    if snapshot_module:
+        try:
+            import importlib
+            import importlib.util
+            if snapshot_module.endswith(".py") or "/" in snapshot_module:
+                # Treat as file path.
+                spec = importlib.util.spec_from_file_location(
+                    "_tui_snapshot_fixture", snapshot_module
+                )
+                if spec is None or spec.loader is None:
+                    raise ImportError(f"cannot load {snapshot_module}")
+                mod = importlib.util.module_from_spec(spec)
+                spec.loader.exec_module(mod)
+            else:
+                mod = importlib.import_module(snapshot_module)
+            snap = getattr(mod, "SNAPSHOT")
+            # v2 dev-only: snapshot modules may export a dict of RuntimeState
+            # field overrides. Lets fixture goldens exercise sort/filter/search/
+            # modal states without adding CLI flags. Spec §5.5.
+            overrides = getattr(mod, "RUNTIME_OVERRIDES", None)
+            if isinstance(overrides, dict):
+                _ALLOWED_OVERRIDES = {
+                    "sort_key", "filter_term", "search_term",
+                    "search_matches", "search_index",
+                    "modal_kind", "modal_scroll", "modal_snap_pending",
+                    "focus_index", "session_scroll",
+                    "input_mode", "input_buffer",
+                    "session_detail_override",
+                }
+                for k, v in overrides.items():
+                    if k in _ALLOWED_OVERRIDES:
+                        setattr(runtime, k, v)
+        except FileNotFoundError as exc:
+            print(f"tui: snapshot module not found: {exc}", file=sys.stderr)
+            return 2
+        except (ImportError, AttributeError) as exc:
+            print(
+                f"tui: failed to load snapshot from {snapshot_module!r}: {exc}",
+                file=sys.stderr,
+            )
+            return 2
+    else:
+        try:
+            snap = _tui_build_snapshot(
+                now_utc=now_utc, skip_sync=True,
+                display_tz_pref_override=getattr(
+                    args, "_display_tz_pref_override", None
+                ),
+            )
+        except Exception:
+            snap = _tui_empty_snapshot(now_utc)
+
+    # --- Render -----------------------------------------------------------
+    # Drift guards run at module import + inside _tui_build_theme itself,
+    # so the explicit inline check has been removed.
+    theme = _tui_build_theme()
+    import io
+    from rich.console import Console
+    # file=StringIO() so console.print() writes into the recording buffer only
+    # (not twice to stdout). export_text() then emits the clean captured copy.
+    console = Console(
+        theme=theme,
+        record=True,
+        width=w,
+        height=h,
+        no_color=not runtime.color_enabled,
+        force_terminal=True,
+        file=io.StringIO(),
+    )
+    bucket = _tui_width_bucket(w)
+    help_panel = _tui_render_help(w, h) if runtime.show_help else None
+    modal_panel = (_tui_render_modal(snap, runtime, w, h)
+                   if runtime.modal_kind else None)
+    # v2: modal wins when both would show. Spec §4.7.
+    overlay = modal_panel or help_panel
+    if runtime.variant == "expressive":
+        frame = _tui_render_variant_b(snap, runtime, w, h, bucket, overlay_panel=overlay)
+    else:
+        frame = _tui_render_variant_a(snap, runtime, w, h, bucket, overlay_panel=overlay)
+    # Layout fills the requested render height with blank rows if the
+    # natural content is shorter, and truncates if taller. Use the
+    # ``_tui_natural_height`` stashed by the variant renderers so the
+    # recorded frame matches the pre-refactor row count (only trailing
+    # whitespace on individual lines is expected to drift; line count
+    # stays identical). Live mode ignores this and renders at terminal
+    # height, which is the desired TUI fill behavior.
+    render_h = getattr(frame, "_tui_natural_height", h) or h
+    console.print(frame, height=render_h)
+    # Default behavior: plain text (matches existing fixture-golden
+    # expectations). FORCE_COLOR=1 opts in to ANSI escapes — used by the
+    # README screenshot pipeline so freeze can render the TUI as a
+    # colored SVG. Goldens never set FORCE_COLOR, so this is byte-safe
+    # for the existing harness.
+    include_styles = os.environ.get("FORCE_COLOR") == "1"
+    sys.stdout.write(console.export_text(styles=include_styles))
+    return 0