@herbertgao/pi-extensions 2026.10.0 → 2026.10.2

package/node_modules/pi-lens/dist/clients/widget-state.js

@@ -1,1808 +0,0 @@
-import { readFileSync } from "node:fs";
-import { stat } from "node:fs/promises";
-import * as path from "node:path";
-import { pathToFileURL } from "node:url";
-import { BoundedFifoMap } from "./bounded-cache.js";
-import { demotePastEofDiagnostics, } from "./diagnostic-line-freshness.js";
-import { visibleWidth } from "./deps/pi-tui.js";
-import { normalizeEphemeralMapKey, normalizeMapKey } from "./path-utils.js";
-import { fitLine } from "./tui-fit.js";
-import { WriteOrderingGuard } from "./write-ordering-guard.js";
-import { collectForwardImportMtimes } from "./blocker-freshness.js";
-import { normalizeMessage } from "./finding-identity.js";
-import { freshnessFromMtime } from "./freshness.js";
-import { PAST_EOF_STALE_MARKER } from "./diagnostic-line-freshness.js";
-import { STALE_LINE_MARKER } from "./stale-marker.js";
-import { anchorsForDiagnostic, applyDispositions, getDisposition, registerWidgetDispositionReconciler, strictAnchorSpan, strictAnchorSpansIn, } from "./diagnostic-dispositions.js";
-import { recordDegradationOnce } from "./degradation-ledger.js";
-/**
- * Canonical key for the `files` map (and `diagnosticsWriteGuard`) — #1020.
- *
- * The SAME file reaches this module under DIFFERENT path forms in one session:
- * forward-slash (`C:/…/x.ts`) from the LSP client + cascade fold via
- * `normalizeFilePath`, and backslash (`C:\…\x.ts`) from mode=full's reconcile
- * writing `result.filePath` and from `path.resolve`/event inputs on Windows.
- * Keyed raw, those coexisted as two entries: `mode=full` re-keyed on read and
- * the clean entry hid the stale one, but `mode=all`'s `formatAllMode` reads the
- * summaries verbatim and rendered the stale `blocking:1` as a 🔴 (#1020) — a
- * resolved state that replayed as still-broken on every `mode=all`.
- *
- * `normalizeEphemeralMapKey` (slash-fold + win32-lowercase, NO filesystem I/O)
- * is chosen over `normalizeMapKey`/`normalizeFilePath`, which call
- * `realpathSync.native()` — real disk I/O on EVERY diagnostic/runner/formatter
- * write, far too heavy for this hot path. The key only needs to be a stable
- * syntactic fold that collapses `\`↔`/` and Windows drive-letter case, which
- * this does; on-disk canonical casing is irrelevant for merely deduplicating a
- * process-local footer cache. The human-readable path is preserved separately
- * on the record's `filePath` (see `toDisplayPath`) for rendering/summaries.
- */
-function fileMapKey(filePath) {
-    return normalizeEphemeralMapKey(filePath);
-}
-/**
- * A diagnostic is "blocking" when pi-lens classifies it as a hard stop
- * (`semantic === "blocking"`). Falls back to severity for sources that
- * don't set `semantic` (raw tsc/eslint diagnostics) so the red dot still
- * fires on traditional compile errors.
- *
- * #1561: exported so the blocker-retire decision in `tools/lsp-diagnostics.ts`
- * asks THIS predicate rather than growing a second severity rule that can drift
- * from what the footer counts.
- */
-export function isBlocking(d) {
-    // #3158 round 2 F1: a finding the agent marked `false-positive`/`suppress` is
-    // not a hard stop — it is an explicitly dismissed one. Before this line only
-    // `countDiagnostics` knew that (it skips tagged entries), so the footer header
-    // counted zero errors while every consumer that asks THIS predicate still
-    // treated the row as blocking: `renderWidget` drew a red ● row for it,
-    // `capStoredDiagnostics` hoisted it ahead of live findings (evicting one at the
-    // display cap), and `getWidgetBlockingFilesForSweep`,
-    // `markWidgetFileBlockersStale` and `reconcileStaleWidgetDependencyBlockers`
-    // admitted it to the turn-end blocker sweep. The mismatch was latent while a
-    // tagged row lived only between a mark and the next scan; retention makes it
-    // durable, so the predicate has to answer for the tag.
-    if (d.disposition)
-        return false;
-    // #1631: a dependency-drift-demoted finding is no longer a hard stop. The gate
-    // sets `stale` rather than dropping the entry (#1419 demote-not-drop), so every
-    // tally and render that asks "is this blocking?" must answer no once demoted.
-    if (d.stale)
-        return false;
-    if (d.semantic === "blocking")
-        return true;
-    if (d.semantic == null && d.severity === "error")
-        return true;
-    return false;
-}
-// ── Module state ─────────────────────────────────────────────────────────────
-const files = new Map();
-let sessionLanguages = [];
-let requestRenderFn = null;
-/**
- * Guards `recordDiagnostics` writes against the same race class fixed for
- * `clients/lsp/client.ts` in #555: pi-lens allows concurrent pipeline runs
- * for the same file across different same-turn edits, so an older edit's
- * (slower) pipeline can finish its `recordDiagnostics` call AFTER a newer
- * edit's (faster) pipeline already recorded fresher diagnostics for that
- * path. Keyed by `filePath`, tokened by `writeIndex` (see
- * `clients/runtime-tool-result.ts:nextWriteIndex`).
- */
-const diagnosticsWriteGuard = new WriteOrderingGuard();
-/**
- * Runner completions also mutate the shared file record: they mark the
- * diagnostics snapshot pending. Keep that mutation in the same per-file order
- * as the final diagnostic replacement, or an older pipeline can set a newer
- * confirmed-clean record back to `(pending)` (#1198).
- */
-const runnerWriteGuard = new WriteOrderingGuard();
-const MAX_STORED_DIAGNOSTICS_PER_FILE = 12;
-/**
- * #3158 / AGENTS.md shape 9: the per-file bound on
- * {@link WidgetDiagnostic.suppressedRetained} rows. Retention is the only rule
- * that lets a record hold findings the latest scan did not report, so it is the
- * only axis of `allDiagnostics` that can grow while the live set shrinks — every
- * other row is replaced wholesale by the next write. Its own constant rather
- * than `MAX_STORED_DIAGNOSTICS_PER_FILE`: that one is the TUI's display cap
- * (`capStoredDiagnostics`) and the two would drift the moment either policy
- * moved. Twelve because the chip is a COUNT, not a list — past a dozen
- * suppressed findings on one unchanged file the exact total stops informing the
- * footer, and the truncation is stated once per file through the ledger rather
- * than dropped silently.
- */
-const MAX_RETAINED_SUPPRESSED_PER_FILE = 12;
-const MAX_INACTIVE_FILE_RECORDS = 1024;
-const ACTIVE_FILE_IDLE_MS = 30 * 60_000;
-export const MAX_LSP_SERVER_RECORDS = 128;
-const lspServers = new BoundedFifoMap(MAX_LSP_SERVER_RECORDS);
-// Pruning is a cold-size-boundary operation. Do not walk the whole file map
-// for every record in a large diagnostics reconciliation; the full-scan path
-// can legitimately create thousands of records in one synchronous batch.
-let nextInactivePruneSize = MAX_INACTIVE_FILE_RECORDS + 1;
-function pruneInactiveFileRecords(now = Date.now()) {
-    if (files.size <= MAX_INACTIVE_FILE_RECORDS)
-        return;
-    const victims = [...files.entries()]
-        .filter(([, rec]) => now - rec.touchedAt > ACTIVE_FILE_IDLE_MS && !hasLiveDiagnostic(rec))
-        .sort(([, a], [, b]) => a.touchedAt - b.touchedAt);
-    for (const [key] of victims) {
-        if (files.size <= MAX_INACTIVE_FILE_RECORDS)
-            break;
-        files.delete(key);
-    }
-}
-function maybePruneInactiveFileRecords() {
-    if (files.size < nextInactivePruneSize)
-        return;
-    pruneInactiveFileRecords();
-    // A live-heavy map may remain above the soft bound. Do not rescan it for
-    // every subsequent file; the next lifecycle starts with a fresh state map.
-    nextInactivePruneSize = Number.POSITIVE_INFINITY;
-}
-// ── Public API ────────────────────────────────────────────────────────────────
-export function setRenderCallback(fn) {
-    requestRenderFn = fn;
-}
-/** Reconcile a mark through the same apply/count/commit path as dispatch. */
-export function reconcileWidgetDisposition(cwd, target, _disposition, content) {
-    const current = getFileDiagnostics(target.filePath);
-    if (!current) {
-        recordDegradationOnce({
-            kind: "widget-disposition-reconcile-fallback",
-            subject: target.filePath,
-            reason: "the marked finding's widget record is absent",
-        });
-        return;
-    }
-    let source = content ?? target.content;
-    if (source === undefined) {
-        try {
-            source = readFileSync(target.filePath, "utf8");
-        }
-        catch {
-            recordDegradationOnce({
-                kind: "widget-disposition-reconcile-fallback",
-                subject: target.filePath,
-                reason: "the disposition anchor content could not be read",
-            });
-            source = "";
-        }
-    }
-    const active = new Set(applyDispositions(current, cwd, target.filePath, source));
-    const normalized = current.flatMap((diagnostic) => {
-        const { strict, weak } = anchorsForDiagnostic(cwd, target.filePath, diagnostic, source);
-        const entry = getDisposition(cwd, strict) ?? getDisposition(cwd, weak);
-        if (!active.has(diagnostic) &&
-            (entry?.disposition === "false-positive" ||
-                entry?.disposition === "suppress")) {
-            // #3183: stamp the strict anchor's own SPAN so the row's lifetime can
-            // later ask the mark instead of the file's mtime. `source` is the content
-            // the anchors above were derived from, so the span is the one the mark
-            // itself binds to. Only the STRICT arm has a span — a weak `suppress`
-            // mark hashes no line — and re-tagging always re-derives it, so a stale
-            // span from an earlier tag can never survive into a new disposition.
-            const { anchorSpan: _previousSpan, ...base } = diagnostic;
-            const span = entry.disposition === "false-positive"
-                ? strictAnchorSpan(source, diagnostic.line)
-                : undefined;
-            return [
-                {
-                    ...base,
-                    disposition: entry.disposition,
-                    flagged: false,
-                    ...(span === undefined ? {} : { anchorSpan: span }),
-                },
-            ];
-        }
-        // #3158 retirement rule 3: a row the store holds ONLY because it was
-        // suppressed (`suppressedRetained`) has no live scan behind it — the last
-        // scan of this file did not report it. Once its mark stops suppressing it
-        // (a strict `false-positive` anchor whose marked line changed), retire it
-        // rather than re-arming it below as a finding the agent would read as
-        // currently present.
-        if (diagnostic.suppressedRetained)
-            return [];
-        const { disposition: _disposition, flagged: _flagged, 
-        // #3183: a row that is no longer disposition-tagged has no mark to
-        // anchor, so its span goes with the tag — leaving it would exempt a LIVE
-        // row from the sweep's mtime gate.
-        anchorSpan: _anchorSpan, ...baseDiagnostic } = diagnostic;
-        return entry?.disposition === "flagged"
-            ? [{ ...baseDiagnostic, flagged: true }]
-            : [baseDiagnostic];
-    });
-    const rec = getOrCreate(target.filePath);
-    const writeIndex = admitWidgetDiagnosticsWrite(target.filePath);
-    if (!diagnosticsWriteGuard.shouldWrite(fileMapKey(target.filePath), writeIndex))
-        return;
-    commitDiagnostics(rec, target.filePath, normalized, Date.now(), fileMapKey(target.filePath), false);
-}
-/** Reserve a widget write for a producer that has no runtime write index. */
-export function admitWidgetDiagnosticsWrite(filePath, writeIndex) {
-    const key = fileMapKey(filePath);
-    const token = writeIndex ?? diagnosticsWriteGuard.nextToken(key);
-    diagnosticsWriteGuard.shouldWrite(key, token);
-    runnerWriteGuard.shouldWrite(key, token);
-    return token;
-}
-registerWidgetDispositionReconciler((cwd, target, disposition) => reconcileWidgetDisposition(cwd, target, disposition, target.content));
-/** Consume the host-side disposition event and refresh the real widget store. */
-export function wireWidgetDispositionSubscriber(args) {
-    args.events?.on?.("pilens:diagnostic:disposition", (data) => {
-        if (!data || typeof data !== "object")
-            return;
-        const payload = data;
-        if (payload.source !== "pi-lens" || typeof payload.cwd !== "string")
-            return;
-        if (typeof payload.filePath !== "string" ||
-            typeof payload.disposition !== "string")
-            return;
-        if (typeof payload.message !== "string")
-            return;
-        if (!["false-positive", "suppress", "defer", "flagged"].includes(payload.disposition))
-            return;
-        reconcileWidgetDisposition(payload.cwd, {
-            cwd: payload.cwd,
-            filePath: payload.filePath,
-            message: typeof payload.message === "string" ? payload.message : "",
-            ...(typeof payload.tool === "string" ? { tool: payload.tool } : {}),
-            ...(typeof payload.rule === "string" ? { rule: payload.rule } : {}),
-            ...(typeof payload.line === "number" ? { line: payload.line } : {}),
-        }, payload.disposition);
-    });
-}
-/**
- * #2275 review F1: files whose dependency-drift-demoted rows the footer has
- * ACTUALLY DRAWN since the last turn end — the delivery population
- * `runtime-turn.ts`'s cap step drains.
- *
- * The footer is not a broadcast of the whole store: `renderWidget` picks ONE
- * record (`withBlocking[0]`) and its first five qualifying entries, and it
- * may not be drawn at all (headless, or a turn with no repaint). Counting a
- * delivery per turn end therefore charged the cap for rows the agent never
- * saw, and retired them a delivery early. A Set (not a counter) is the right
- * shape: the widget repaints many times per turn, and all of those repaints
- * are the SAME one delivery.
- */
-const renderedDependencyDriftFiles = new Set();
-export function clearWidgetState() {
-    files.clear();
-    renderedDependencyDriftFiles.clear();
-    lspServers.clear();
-    sessionLanguages = [];
-    requestRenderFn = null;
-    diagnosticsWriteGuard.clear();
-    runnerWriteGuard.clear();
-    nextInactivePruneSize = MAX_INACTIVE_FILE_RECORDS + 1;
-}
-// v1 → v2 (#1186): per-entry `WidgetDiagnostic.observedAt`. v2 is a SUPERSET of
-// v1 (the field is additive/optional), so `importWidgetState` accepts either and
-// migrates a v1 record by inheriting each entry's `observedAt` from the record's
-// `touchedAt`. A v1 file must never be rejected (that would silently drop resume
-// diagnostics) nor crash.
-//
-// v2 → v3 (#1631 review F3): a dependency-drift `stale` demotion is a verdict
-// about THIS session's disk state — the next session starts with a fresh (and
-// possibly different) dependency graph, so a demotion from the last session
-// must not outlive it. `exportWidgetState` had been serializing `stale: true`
-// verbatim; on restore, a genuine type error stayed non-blocking across a
-// resume with the dependency pinned back into the past. Same shape as the
-// #1348 failed-formatter precedent one field over: `importWidgetState` strips
-// `stale` on every restored entry so a resumed session re-evaluates from
-// scratch instead of inheriting a possibly-stale-itself verdict.
-export const WIDGET_STATE_VERSION = 3;
-/**
- * Snapshot the per-file widget diagnostics for persistence (#190). Excludes
- * `lspServers` — those are process-bound (servers re-spawn fresh on the next
- * launch), so restoring their "ready" status would be misleading.
- */
-export function exportWidgetState() {
-    return {
-        version: WIDGET_STATE_VERSION,
-        sessionLanguages: [...sessionLanguages],
-        files: [...files.values()].map((rec) => ({
-            filePath: rec.filePath,
-            runners: [...rec.runners.entries()],
-            formatters: [...rec.formatters.entries()],
-            diagnostics: rec.diagnostics,
-            allDiagnostics: rec.allDiagnostics,
-            diagnosticCounts: rec.diagnosticCounts,
-            hasFinalDiagnosticsSnapshot: rec.hasFinalDiagnosticsSnapshot,
-            touchedAt: rec.touchedAt,
-        })),
-    };
-}
-/**
- * Restore a {@link PersistedWidgetState} snapshot (#190 resume rehydration).
- * Replaces the in-memory `files` map; ignores snapshots from a different
- * version. Triggers a re-render if a callback is registered.
- */
-/**
- * #1186 v1→v2 migration: stamp each entry that lacks a per-entry `observedAt`
- * with the record's `touchedAt` (the single stamp the whole record shared under
- * v1). Non-mutating; entries that already carry a stamp (a v2 record) pass
- * through untouched.
- *
- * #1631 review F3 (v2→v3): also strips `stale` from every entry. A dependency
- * drift demotion is scoped to the session that observed the drift; restoring it
- * verbatim would let a real blocking finding resume as permanently non-blocking
- * (#1348 precedent — see `WIDGET_STATE_VERSION`'s doc comment).
- */
-function migrateEntryStamps(entries, recordTouchedAt) {
-    return (entries ?? []).map((d) => {
-        // #2275: strip staleDeliveryCount and footerRetired alongside stale — a
-        // delivery count (and the footer-hide it earned) without the demotion
-        // it counts is meaningless, and a resumed session re-evaluates
-        // dependency-drift from scratch (#1631 review F3) same as the `stale`
-        // bit itself. A restored footer has shown the row zero times.
-        // Fix-round 3 (#2275 review F2): `staleReason` left the round trip
-        // unstripped, so a resumed row could carry `staleReason:
-        // "dependency-drift"` with `stale` gone — `isDependencyDriftDemoted`
-        // reads `stale === true`, so that combination reads as clean today, but
-        // the stray reason is still a demotion label with no demotion behind
-        // it, the same "meaningless without the bit it counts" shape as the
-        // other three fields. Strip it alongside them.
-        const { stale: _stale, staleDeliveryCount: _staleDeliveryCount, footerRetired: _footerRetired, staleReason: _staleReason, ...rest } = d;
-        return rest.observedAt == null
-            ? { ...rest, observedAt: recordTouchedAt }
-            : rest;
-    });
-}
-export function importWidgetState(state) {
-    // Accept any known-or-older version and migrate (#1186): reject a missing
-    // snapshot, a missing/non-numeric `version` (NaN/undefined/null — the
-    // pre-#1186 guard `version !== WIDGET_STATE_VERSION` rejected these, and
-    // loosening that would silently admit a malformed/foreign snapshot), or a
-    // FUTURE version this build can't understand. Rejecting a v1
-    // (pre-per-entry-stamp) file, by contrast, would silently drop all resume
-    // diagnostics — so v1..current are accepted and migrated.
-    if (!state ||
-        typeof state.version !== "number" ||
-        state.version < 1 ||
-        state.version > WIDGET_STATE_VERSION) {
-        return false;
-    }
-    files.clear();
-    // A resumed session's writeIndex counter starts fresh (#190 rehydration is
-    // process-bound like lspServers, see the export above) — any ordering
-    // tokens tracked before the restore no longer correspond to anything, so
-    // drop them rather than risk a legitimate post-resume write being read as
-    // "superseded" against a stale token.
-    diagnosticsWriteGuard.clear();
-    runnerWriteGuard.clear();
-    for (const f of state.files ?? []) {
-        // Fold persisted keys through the same normalizer as live writes (#1020),
-        // or a persisted forward-slash key stays split from a fresh backslash key
-        // across a resumed session — a primary repro condition. Keep a readable
-        // display path on the record.
-        // #1186 migration: a v1 record's entries have no per-entry `observedAt`.
-        // Inherit the record's `touchedAt` (a safe, over-conservative default —
-        // the whole record shared that one stamp before), so the per-entry stale
-        // gate has a concrete observation time and never treats `undefined` as
-        // epoch-0 (which would drop every migrated entry on the first sweep).
-        const recordTouchedAt = f.touchedAt ?? Date.now();
-        // #1631 review V1: `diagnosticCounts` is DERIVED from `allDiagnostics`
-        // everywhere else in this module (see the `countDiagnostics(rec.allDiagnostics)`
-        // call sites) — it must be recomputed here too, from the just-migrated
-        // (stale-stripped) entries, not restored verbatim from the snapshot. The
-        // persisted counts were computed while a finding was still demoted
-        // (`blocking: 0`); once F3 strips `stale` so `isBlocking` reports true
-        // again, a verbatim count would still say `blocking: 0` for a record
-        // whose entries ARE blocking — inverting the one-predicate invariant
-        // `isBlocking` exists to hold. Every consumer (`getFileDiagnosticSummaries`,
-        // the record-tier classifier, the footer) trusts `diagnosticCounts`, not a
-        // live re-scan of the entries, so a derived count out of sync with its own
-        // entries is silently wrong everywhere at once.
-        const migratedAllDiagnostics = migrateEntryStamps(f.allDiagnostics, recordTouchedAt);
-        files.set(fileMapKey(f.filePath), {
-            filePath: f.filePath,
-            runners: new Map(f.runners ?? []),
-            // Failure entries do NOT survive a session restore (#1348 review):
-            // a fmt-failed marker is live advice about THIS session's last
-            // attempt; rehydrating one from a snapshot shows a stale failure the
-            // current session never observed (and same-mtime fixes would never
-            // clear it). Successes rehydrate as before.
-            formatters: new Map((f.formatters ?? []).filter(([, outcome]) => outcome?.success !== false)),
-            diagnostics: migrateEntryStamps(f.diagnostics, recordTouchedAt),
-            allDiagnostics: migratedAllDiagnostics,
-            diagnosticCounts: countDiagnostics(migratedAllDiagnostics),
-            hasFinalDiagnosticsSnapshot: f.hasFinalDiagnosticsSnapshot ?? false,
-            touchedAt: recordTouchedAt,
-        });
-    }
-    pruneInactiveFileRecords();
-    sessionLanguages = state.sessionLanguages ?? [];
-    requestRenderFn?.();
-    return true;
-}
-export function setSessionLanguages(langs) {
-    sessionLanguages = langs;
-    requestRender();
-}
-/** File-kinds detected in use this session (#170 staleness scope). */
-export function getSessionLanguages() {
-    return [...sessionLanguages];
-}
-/**
- * Distinct serverIds with a failed spawn record (#170). Raw — the per-language
- * coverage check (a live sibling) and the in-use staleness filter live in
- * `selectLspStatus`, which joins this against the alive set and session kinds.
- */
-export function getFailedLspServerIds() {
-    const ids = [];
-    const seen = new Set();
-    for (const rec of lspServers.values()) {
-        if (rec.status !== "failed" || seen.has(rec.serverId))
-            continue;
-        seen.add(rec.serverId);
-        ids.push(rec.serverId);
-    }
-    return ids;
-}
-export function recordFormatter(filePath, formatter, changed, success, outcome) {
-    const rec = getOrCreate(filePath);
-    rec.formatters.set(formatter, { changed, success, outcome });
-    rec.touchedAt = Date.now();
-    files.set(fileMapKey(filePath), rec);
-    requestRender();
-}
-export function recordRunner(filePath, runnerId, status, diagnosticCount, durationMs, writeIndex) {
-    const key = fileMapKey(filePath);
-    if (!runnerWriteGuard.shouldWrite(key, writeIndex))
-        return;
-    // Advance the final-diagnostics guard too. A runner completion is part of
-    // this pipeline's ordered write, even though its immediate effect is only to
-    // mark the record pending. This prevents an older final replacement from
-    // landing after a newer runner state (and vice versa).
-    diagnosticsWriteGuard.shouldWrite(key, writeIndex);
-    const rec = getOrCreate(filePath);
-    rec.runners.set(runnerId, { status, count: diagnosticCount, durationMs });
-    rec.hasFinalDiagnosticsSnapshot = false;
-    rec.touchedAt = Date.now();
-    files.set(fileMapKey(filePath), rec);
-    requestRender();
-}
-/**
- * Collapse a (possibly multi-line) diagnostic message to a single line.
- * TS2769 / "no overload matches" and many compiler errors are multi-line;
- * embedded newlines/tabs would otherwise render across several widget rows
- * (and break the `L<line>: <message>` inline-blocker format), so flatten all
- * whitespace runs to a single space before storing.
- */
-function toSingleLineMessage(message) {
-    return (message ?? "").replace(/\s+/g, " ").trim();
-}
-/** Build the OSC-8 target used by every stored diagnostic row. */
-export function widgetDiagnosticUri(filePath, line, column) {
-    const base = pathToFileURL(filePath).href;
-    return line != null
-        ? `${base}#L${line}${column != null ? `:${column}` : ""}`
-        : base;
-}
-export function recordDiagnostics(filePath, diagnostics, writeIndex, 
-// #1093: when the truth was OBSERVED, not when it's being written. Defaults
-// to `Date.now()` for the per-edit/live path (observed now). A reconcile
-// replaying a CACHED view (e.g. the workspace-diagnostics cache-hit branch in
-// `tools/lsp-diagnostics.ts`) must pass the cache entry's own scan timestamp
-// here — otherwise a repeat "fresh check" that merely re-serves a stale
-// cached view keeps bumping `touchedAt` to now(), permanently disarming
-// `reconcileStaleWidgetFiles`'s `mtimeMs > touchedAt` gate so a resolved
-// finding renders forever (the #1092 touchedAt-re-arming defect).
-observedAt) {
-    // Drop a write that's superseded by a later same-turn edit to this file
-    // whose pipeline finished first (same race class as #555). No cache write,
-    // no count/timestamp update, no render trigger — the recorded state must
-    // stay exactly as the fresher write left it. `writeIndex` omitted (e.g.
-    // the `clients/mcp/analyze.ts` on-demand call site, which has no per-edit
-    // ordering token) always proceeds, same as version-less LSP servers in the
-    // #555 guard.
-    const key = fileMapKey(filePath);
-    if (!diagnosticsWriteGuard.shouldWrite(key, writeIndex))
-        return false;
-    // Keep runner state ordered with the final diagnostic replacement. The
-    // guards are deliberately advanced in both directions because either verb
-    // may be the first completion from a pipeline.
-    runnerWriteGuard.shouldWrite(key, writeIndex);
-    // Resolve the observation time ONCE (#1186): every incoming entry is stamped
-    // with it, and it also seeds the record's `touchedAt`. A fresh write (no
-    // `observedAt`) is observed now.
-    const observedTs = observedAt ?? Date.now();
-    const rec = getOrCreate(filePath, key);
-    commitDiagnostics(rec, filePath, normalizeDiagnostics(filePath, diagnostics, observedTs), observedTs, key);
-    return true;
-}
-/** Map the raw diagnostic shape callers pass into stored {@link WidgetDiagnostic}s.
- * Every produced entry is stamped with `observedTs` (#1186) — the time THIS batch
- * of diagnostics was observed — so the per-entry stale gate can later drop just
- * the entries older than the file's mtime rather than the whole record. */
-function normalizeDiagnostics(filePath, diagnostics, observedTs) {
-    return diagnostics.map((d) => {
-        const rule = d.rule ?? d.id;
-        return {
-            severity: d.severity ?? "info",
-            semantic: d.semantic,
-            message: toSingleLineMessage(d.message),
-            line: d.line,
-            col: d.column,
-            rule,
-            tool: d.tool,
-            uri: widgetDiagnosticUri(filePath, d.line, d.column),
-            observedAt: observedTs,
-        };
-    });
-}
-/**
- * The identity a retained suppressed row is matched against a later scan's
- * entries by (#3158), so a scan that reports the finding AGAIN replaces the
- * retained row instead of standing beside it as a duplicate.
- *
- * These are the WEAK ANCHOR's identity parts — `computeWeakAnchor`
- * (`clients/diagnostic-dispositions.ts`) hashes `[tool, rule,
- * normalizeMessage(message)]` over the file — deliberately, because the weak
- * anchor is how the disposition store itself decides which findings a mark
- * covers. Matching at any finer granularity means the store can disagree with
- * the mark that created the row.
- *
- * Round 2 F2: `line` USED to be part of it, and that was the finer granularity
- * the paragraph above forbids. A writer that re-reports the marked finding one
- * line down — `clients/pipeline.ts`'s per-edit `recordDiagnostics`, or
- * `reconcileCascadeNeighborLspErrors`, neither of which runs
- * `reconcileStaleWidgetFiles` — left the retained row standing beside the live
- * one, so a single finding was counted live AND in `suppressed: N`.
- *
- * `computeWeakAnchor` itself is not called here: it needs `cwd` to build its
- * path component, which `commitDiagnostics` does not have and could only get by
- * widening `recordDiagnostics`, `reconcileScanDiagnostics` and
- * `reconcileCorrelatedScanDiagnostics` up through `tools/`. That component is
- * constant across one file record — both sides of this comparison are the same
- * file — so it would discriminate nothing. `normalizeMessage` is the shared
- * derivation from `clients/finding-identity.ts` that the anchor uses, imported
- * rather than re-spelled.
- *
- * #3183: this identity is OCCURRENCE-BLIND, and for a STRICT (`false-positive`)
- * row that is not enough on its own. Two findings of one rule with identical
- * messages on different lines collapse onto it, so a scan reporting only the
- * UNMARKED one matched the marked row and replaced it — the chip lost a mark
- * that was still applying. A collision is therefore resolved by the second half
- * of {@link WidgetDiagnostic.anchorSpan}'s contract, in `retainSuppressedRows`:
- * only a colliding incoming finding sitting ON the marked line is the same
- * finding. That test, not this one, is what keeps the round-2 F2 case above
- * closed — the marked line's TEXT moved, so its span still matches.
- */
-function retentionIdentity(d) {
-    return JSON.stringify([
-        d.tool ?? "",
-        d.rule ?? "",
-        normalizeMessage(d.message),
-    ]);
-}
-/** `filePath`'s current content, or `undefined` when it cannot be read. */
-function readContentOrUndefined(filePath) {
-    try {
-        return readFileSync(filePath, "utf8");
-    }
-    catch {
-        // Only reached once this module has established there IS a span-carrying row
-        // to ask about, so the record is bounded by that population — and worth
-        // making: with no content both callers fall back to the pre-#3183 rules (the
-        // per-entry mtime gate, the coarse retention identity), which is the
-        // conservative direction but can retire a mark that still stands.
-        recordDegradationOnce({
-            kind: "widget-mark-anchor-unreadable",
-            subject: filePath,
-            reason: "the marked file could not be read, so its suppressed rows fall back to the file-mtime retirement gate and the footer chip may under-count it",
-        });
-        return undefined;
-    }
-}
-/**
- * {@link strictAnchorSpansIn} over `filePath`'s current content (#3183), or
- * `undefined` when it cannot be read.
- *
- * Synchronous in both callers, matching `applyCachedDispositions`
- * (`tools/lens-diagnostics.ts`) — the seam that asks the very same question on
- * the very same read path, with `statSync`/`readFileSync`. An async read here
- * would add two unbounded awaits to a module a registered hook handler imports
- * directly (`tests/config/hook-await-bounds.test.ts`) for no behavioural gain.
- */
-function readAnchorSpans(filePath) {
-    const content = readContentOrUndefined(filePath);
-    return content === undefined ? undefined : strictAnchorSpansIn(content);
-}
-/**
- * #3158: carry the record's disposition-tagged rows across a whole-replace
- * write when the incoming set no longer reports them, so a `false-positive` /
- * `suppress` mark keeps contributing to the footer's `suppressed: N` chip. See
- * {@link WidgetDiagnostic.suppressedRetained} for the full lifetime statement.
- *
- * Retained rows are APPENDED after the incoming ones: `capStoredDiagnostics`
- * fills the TUI's display list from the front, so a retained row never evicts a
- * live finding from the rendered set.
- *
- * The tagged-row check comes FIRST so a project with no disposition marks — the
- * overwhelmingly common case, and this runs on the per-edit write path — pays
- * one predicate per stored row and never builds the identity set.
- */
-function retainSuppressedRows(previous, incoming, filePath) {
-    const tagged = previous.filter((d) => d.disposition !== undefined);
-    if (tagged.length === 0)
-        return incoming;
-    const reported = new Map();
-    for (const d of incoming) {
-        const id = retentionIdentity(d);
-        const bucket = reported.get(id);
-        if (bucket)
-            bucket.push(d);
-        else
-            reported.set(id, [d]);
-    }
-    // #3183: the ONE file read on this path, and only for the state the coarse
-    // identity genuinely cannot resolve — a strict-anchored row whose identity an
-    // incoming finding collides with. Every other write (no marks at all, a mark
-    // with no collision, a weak `suppress` mark) stays zero-I/O, which is what
-    // keeps this off the per-edit `recordDiagnostics` cost.
-    const content = tagged.some((d) => d.anchorSpan !== undefined && reported.has(retentionIdentity(d)))
-        ? readContentOrUndefined(filePath)
-        : undefined;
-    const spans = content === undefined ? undefined : strictAnchorSpansIn(content);
-    const retained = tagged.filter((d) => {
-        const collisions = reported.get(retentionIdentity(d));
-        // Nothing re-reported this finding, so retention rule 2 has nothing to say.
-        if (collisions === undefined)
-            return true;
-        // No span to ask (a weak `suppress` mark, or a row from an older build), or
-        // the file is unreadable: fall back to the pre-#3183 coarse rule. The safe
-        // direction is to UNDER-count, never to leave a phantom duplicate standing
-        // beside a live finding.
-        if (d.anchorSpan === undefined || spans === undefined)
-            return false;
-        // The mark stopped applying — its line is gone from the file — so the
-        // colliding finding IS this one, re-reported now that the filter no longer
-        // drops it (retirement rule 2, and the same verdict the sweep reaches).
-        if (!spans.has(d.anchorSpan))
-            return false;
-        // The mark still applies. Only a collision sitting ON the marked line is
-        // the same finding (its line number may have shifted under an edit — #3158
-        // round 2 F2); any other is a DIFFERENT occurrence this strict anchor never
-        // covered, and replacing the row with it loses a live mark (#3183).
-        return !collisions.some((i) => strictAnchorSpan(content, i.line) === d.anchorSpan);
-    });
-    if (retained.length > MAX_RETAINED_SUPPRESSED_PER_FILE) {
-        // Freshest observations win the cap. One record per FILE, never one per
-        // dropped row (AGENTS.md "bounded observability"): the chip under-counts
-        // by exactly `dropped` for this file until its content changes.
-        retained.sort((a, b) => (b.observedAt ?? 0) - (a.observedAt ?? 0));
-        recordDegradationOnce({
-            kind: "widget-suppressed-retention-capped",
-            subject: filePath,
-            reason: `${retained.length - MAX_RETAINED_SUPPRESSED_PER_FILE} suppressed row(s) beyond the ${MAX_RETAINED_SUPPRESSED_PER_FILE}-row per-file retention cap are no longer counted by the footer chip`,
-        });
-        retained.length = MAX_RETAINED_SUPPRESSED_PER_FILE;
-    }
-    return [
-        ...incoming,
-        ...retained.map((d) => d.suppressedRetained ? d : { ...d, suppressedRetained: true }),
-    ];
-}
-/** Store `normalized` as the record's complete diagnostic set: recompute counts,
- * cap the display list, stamp `touchedAt` (at `observedAt` when given, else now
- * — #1093), persist, and re-render. The caller decides what `normalized`
- * contains (a full replace, or a merge — see `reconcileCascadeNeighborLspErrors`). */
-function commitDiagnostics(rec, filePath, incoming, observedAt, key = fileMapKey(filePath), 
-/**
- * #3158: apply the suppressed-row retention rule to this write. True for
- * every seam that hands in a fresh SCAN of the file (`recordDiagnostics` and
- * both reconciles), false for `reconcileWidgetDisposition`, which re-maps the
- * record's own current rows and is the seam that RETIRES a retained row —
- * re-adding what it just dropped would make retirement rule 3 inert.
- */
-retainSuppressed = true) {
-    const normalized = retainSuppressed
-        ? retainSuppressedRows(rec.allDiagnostics, incoming, filePath)
-        : incoming;
-    rec.diagnosticCounts = countDiagnostics(normalized);
-    rec.diagnostics = capStoredDiagnostics(normalized);
-    rec.allDiagnostics = normalized;
-    rec.hasFinalDiagnosticsSnapshot = true;
-    // Record-level `touchedAt` is the FRESHEST per-entry observation in the merged
-    // set (#1186) — drives render recency and the empty-record stale gate. On a
-    // merge (`reconcileCascadeNeighborLspErrors`) this is the newest of the
-    // preserved + incoming entries, not the (possibly aging) incoming stamp, so a
-    // record holding a fresh preserved entry doesn't sort/gate as stale. Empty set
-    // falls back to the passed `observedAt` (or now).
-    rec.touchedAt = freshestObservation(normalized, observedAt ?? Date.now());
-    files.set(key, rec);
-    requestRender();
-}
-export function classifyDiagnosticTier(d) {
-    if (d.severity === "error")
-        return "error";
-    if (d.severity === "warning")
-        return "warning";
-    if (d.severity === "hint" || d.severity === "info")
-        return "advisory";
-    return undefined;
-}
-/** Recompute the {blocking, errors, warnings} tally for a diagnostic set.
- * `hint`/`info` tier findings are tallied separately — see
- * {@link classifyDiagnosticTier} and {@link countAdvisories} — and never
- * inflate `warnings` (#2414). */
-function countDiagnostics(diags) {
-    let blocking = 0;
-    let errors = 0;
-    let warnings = 0;
-    for (const diagnostic of diags) {
-        if (diagnostic.disposition)
-            continue;
-        if (isBlocking(diagnostic))
-            blocking++;
-        // A past-EOF stale entry keeps its severity for display purposes but is
-        // excluded from the error/warning tallies alongside blocking — its cited
-        // coordinate is no longer trustworthy, same reasoning as `isBlocking`.
-        // A dependency-drift demotion (#1631 criterion 3) is different: the
-        // finding itself is still real evidence, only its BLOCKING authority is
-        // revoked until re-confirmed, so it stays in the error/warning tally.
-        if (diagnostic.stale &&
-            (diagnostic.staleReason ?? "past-eof") === "past-eof")
-            continue;
-        const tier = classifyDiagnosticTier(diagnostic);
-        if (tier === "error")
-            errors++;
-        else if (tier === "warning")
-            warnings++;
-        // tier === "advisory" (hint/info): intentionally excluded from the
-        // footer's warning tally (#2414) — see `countAdvisories` for the
-        // parallel count `getFileDiagnosticSummaries` exposes to
-        // `lens_diagnostics` so a hint-only file still shows up as "has issues".
-    }
-    return { blocking, errors, warnings };
-}
-/** Count `hint`/`info` tier findings using the same past-EOF exclusion as
- * {@link countDiagnostics} (#2414). Kept as a standalone accessor rather than
- * a new `FileRecord.diagnosticCounts` field — the footer never renders this
- * number, only `getFileDiagnosticSummaries` (mode=all) needs it, to keep a
- * hint-only file from being silently dropped from a detailed listing. */
-function countAdvisories(diags) {
-    let advisories = 0;
-    for (const diagnostic of diags) {
-        if (diagnostic.stale &&
-            (diagnostic.staleReason ?? "past-eof") === "past-eof")
-            continue;
-        if (classifyDiagnosticTier(diagnostic) === "advisory")
-            advisories++;
-    }
-    return advisories;
-}
-/**
- * Apply the #1641 past-EOF gate to `rec` in place: demote (never drop) any
- * stored diagnostic whose cited line exceeds the file's CURRENT on-disk line
- * count, then recompute the capped/full diagnostic lists and the counts so
- * every reader (the TUI render loop, `getFileDiagnosticSummaries`,
- * `getFileDiagnostics`) sees one consistent, already-gated record instead of
- * each re-deriving its own verdict. Cheap on the common case: one memoized
- * stat per record (see `getCachedLineCount`), a full recount only when the
- * file's mtime moved since the last check.
- *
- * No `resync` callback here deliberately: this gate runs on the TUI's render
- * path and the bus-publish/mark-tool read accessors, all of which fire far
- * more often than an agent's own edit/tool cadence. Triggering a document
- * resync from every one of those reads would storm the LSP with didOpen calls
- * for a file whose drift hasn't yet been fixed. `tools/lens-diagnostics.ts`'s
- * `formatAllMode` — an explicit, agent-invoked tool call — is the resync
- * trigger point (#1641 criterion 2); this gate still demotes/logs on its own.
- *
- * The gate RE-DERIVES every entry's `stale` flag on every call — it is never
- * a one-way latch. A transient shrink that later restores (truncate-then-
- * write, a formatter pass, a checkout) un-demotes on its own next read, so
- * this always persists the freshly-derived array back into the record, not
- * only when `demotedCount` (which counts RISING edges only, for telemetry)
- * is nonzero.
- */
-function applyPastEofGate(rec, lineCountCache) {
-    if (rec.allDiagnostics.length === 0)
-        return;
-    const { diagnostics } = demotePastEofDiagnostics({
-        store: "widget-state",
-        cwd: process.cwd(),
-        filePath: rec.filePath,
-        diagnostics: rec.allDiagnostics,
-        lineCountCache,
-    });
-    rec.allDiagnostics = diagnostics;
-    rec.diagnostics = capStoredDiagnostics(diagnostics);
-    rec.diagnosticCounts = countDiagnostics(diagnostics);
-}
-/** The newest per-entry `observedAt` in `diags`, or `fallback` when empty (or no
- * entry carries a stamp). Used as the record-level `touchedAt` (#1186). */
-function freshestObservation(diags, fallback) {
-    let newest;
-    for (const d of diags) {
-        if (d.observedAt != null &&
-            (newest === undefined || d.observedAt > newest)) {
-            newest = d.observedAt;
-        }
-    }
-    return newest ?? fallback;
-}
-/**
- * A stored diagnostic that came from a language SERVER (tsserver, pyright, …)
- * AND is an error/blocking finding. `convertLspDiagnostics` tags every
- * language-server diagnostic with `tool: "lsp"` and `retagAuxiliaryDiagnostics`
- * re-tags auxiliary-LSP findings (opengrep/ast-grep/zizmor/typos) to their real
- * tool id, so `tool === "lsp"` uniquely identifies a genuine language-server
- * entry. Used by the cascade merge below to decide which existing entries an
- * errors-only LSP re-check is entitled to replace.
- */
-function isLspErrorEntry(d) {
-    return (d.tool === "lsp" && (d.semantic === "blocking" || d.severity === "error"));
-}
-/**
- * Reconcile a CONFIRMED cascade neighbor re-check (#1093) into the footer,
- * MERGING rather than whole-replacing (unlike `recordDiagnostics`).
- *
- * The cascade only re-checks a neighbor through its LANGUAGE SERVER, and only
- * for ERRORS (`severity === 1`; see `clients/dispatch/integration.ts`). A plain
- * full-replace would therefore erase the neighbor's live findings from OTHER
- * sources that the cascade never re-examined — biome/ruff/ast-grep runner
- * findings, and even the language server's own WARNINGS — turning a
- * cross-file-error re-check into a silent false-clean for everything else
- * (#533). So we replace ONLY the existing LSP-error entries
- * (`isLspErrorEntry`) with the cascade's fresh LSP errors and preserve
- * everything else verbatim. A stale LSP warning or biome finding therefore
- * survives an errors-only cascade — correct, because this check never looked at
- * it; it self-corrects on the next per-edit dispatch or a `lens_diagnostics`
- * scan (which DO re-examine every source).
- *
- * Only ever call this for a CONFIRMED result (a valid passive snapshot or a
- * completed, NON-inconclusive active touch — #571). `writeIndex` and
- * `observedAt` behave exactly as in `recordDiagnostics`.
- */
-export function reconcileCascadeNeighborLspErrors(filePath, lspErrorDiagnostics, writeIndex, observedAt) {
-    const key = fileMapKey(filePath);
-    if (!diagnosticsWriteGuard.shouldWrite(key, writeIndex))
-        return;
-    runnerWriteGuard.shouldWrite(key, writeIndex);
-    // #1186: the INCOMING LSP errors are stamped at THIS observation time
-    // (`observedAt`, e.g. an aging passive snapshot's `entry.ts`, or now for a
-    // fresh active touch). The PRESERVED entries keep their OWN prior per-entry
-    // `observedAt` — a fresh per-edit finding preserved through this errors-only
-    // merge is NOT re-aged to the incoming stamp. That per-entry split is exactly
-    // what lets `reconcileStaleWidgetFiles` drop the stale incoming entry while
-    // keeping the newer preserved one, instead of dropping the whole record.
-    const observedTs = observedAt ?? Date.now();
-    const rec = getOrCreate(filePath);
-    const incoming = normalizeDiagnostics(filePath, lspErrorDiagnostics, observedTs);
-    const preserved = rec.allDiagnostics.filter((d) => !isLspErrorEntry(d));
-    commitDiagnostics(rec, filePath, [...incoming, ...preserved], observedTs);
-}
-/**
- * Reconcile a diagnostics result obtained OUTSIDE the per-edit dispatch
- * pipeline — a `lens_diagnostics` mode=full workspace scan, or a standalone
- * `lsp_diagnostics` on-demand check — into the footer cache (#571).
- *
- * `recordDiagnostics` is otherwise only reachable from `pipeline.ts`'s
- * per-edit dispatch, so a file that becomes stale/fresh purely because of a
- * change to some OTHER file it depends on (and is never itself re-edited
- * through pi-lens) has no path to correct the footer — a full scan proves
- * the fresher truth but had nowhere to put it. This is that path, shared by
- * both call sites so there's exactly one place that decides whether a scan
- * result is trustworthy enough to write.
- *
- * `confirmed` MUST be false for any result the caller can't vouch for — a
- * timed-out/inconclusive LSP check (see #570) must never present as
- * "confirmed clean" in the footer, and must not clobber a real prior
- * confirmed-dirty entry either. Non-confirmed results are silently skipped,
- * leaving whatever the footer already had (stale-but-real beats
- * fresh-but-fabricated).
- *
- * `writeIndex` should be a freshly-drawn token from the same monotonic
- * source the per-edit pipeline uses (`RuntimeCoordinator.nextWriteIndex()`)
- * so `recordDiagnostics`'s existing `WriteOrderingGuard` (#555) can tell a
- * scan-originated write apart from a concurrent, genuinely newer per-edit
- * write for the same file — an omitted `writeIndex` always proceeds (same
- * version-less fallback `recordDiagnostics` already documents), which is
- * only safe for callers with no ordering token to give (e.g. tests).
- *
- * `observedAt` (#1093) is the wall-clock time the diagnostics were actually
- * OBSERVED — pass it whenever the reconciled result is a replay of an older
- * CACHED observation (the workspace-diagnostics cache-hit branch), so
- * `touchedAt` records when the truth was seen, not when it was written. Omit
- * it for genuinely fresh observations (a just-completed touch/scan), which are
- * observed now.
- */
-export function reconcileScanDiagnostics(filePath, diagnostics, confirmed, writeIndex, observedAt) {
-    if (!confirmed)
-        return false;
-    return recordDiagnostics(filePath, diagnostics, writeIndex, observedAt);
-}
-/**
- * Commit the already-correlated result of a full diagnostics scan.
- *
- * Unlike {@link reconcileScanDiagnostics}, this seam accepts widget diagnostics
- * that have already been merged across producing lanes. `lens_diagnostics`
- * uses it after correlating the confirmed LSP sweep, the cheap project scan,
- * and retained widget rows. This prevents a broken LSP lane from hiding
- * independently-produced `ast-grep-napi` findings in the widget count (#1888).
- * Existing per-entry observation times survive the merge; newly-added project
- * rows inherit the scan's observation time.
- */
-export function reconcileCorrelatedScanDiagnostics(filePath, diagnostics, writeIndex, observedAt) {
-    const key = fileMapKey(filePath);
-    if (!diagnosticsWriteGuard.shouldWrite(key, writeIndex))
-        return;
-    runnerWriteGuard.shouldWrite(key, writeIndex);
-    const observedTs = observedAt ?? Date.now();
-    const correlated = diagnostics.map((diagnostic) => ({
-        ...diagnostic,
-        observedAt: diagnostic.observedAt ?? observedTs,
-    }));
-    commitDiagnostics(getOrCreate(filePath, key), filePath, correlated, observedTs, key);
-}
-/**
- * Drop widget entries whose file changed on disk after pi-lens last recorded
- * them (`mtimeMs > touchedAt` → the recorded diagnostics predate the current
- * content → stale) or that no longer exist. Keeps `lens_diagnostics` from
- * surfacing findings the agent already fixed (or that an external edit
- * invalidated). Async with concurrent stats — call on read, never on the typing
- * path. Returns how many entries were dropped (so callers can tell the agent
- * those files changed and need a `mode=full` rescan rather than reading as
- * clean).
- */
-export async function reconcileStaleWidgetFiles() {
-    const entries = [...files.entries()];
-    const verdicts = await Promise.all(
-    // `mapKey` is the normalized `files` key (used for deletion); stat the
-    // record's real display path, not the lowercased key (#1020).
-    entries.map(async ([mapKey, rec]) => {
-        let mtimeMs;
-        try {
-            mtimeMs = (await stat(rec.filePath)).mtimeMs;
-        }
-        catch {
-            return { mapKey, action: "drop" }; // deleted / unreadable → drop
-        }
-        // A clean record (no findings) has no per-entry stamps to consult —
-        // gate it on the record's own `touchedAt` exactly as before, so a ✓
-        // entry for a file that changed on disk still drops.
-        //
-        // #1631 review F11 follow-up: a THIRD site hitting the same Windows
-        // mtime-vs-`Date.now()` skew F2 named for `blocker-freshness.ts` and
-        // `reconcileStaleWidgetDependencyBlockers` — a file's mtime can lead
-        // `Date.now()` by up to ~11.4ms measured (#1491/#1498 precedent), so a
-        // record touched immediately after its own write dropped here at +1ms
-        // with zero real drift. Found investigating F11: `lens-diagnostics-
-        // mode-all-freshness.test.ts`'s "control" case (and its sibling) failed
-        // under this gate, not from cross-test state. Same tolerance, same
-        // constant, for the same reason.
-        if (rec.allDiagnostics.length === 0) {
-            return freshnessFromMtime({ mtimeMs, referenceMs: rec.touchedAt })
-                .verdict === "stale"
-                ? { mapKey, action: "drop" }
-                : { mapKey, action: "keep" };
-        }
-        // #1186 per-ENTRY gate: drop only the entries observed BEFORE the file's
-        // current mtime; keep the rest. A merged record can hold a fresh
-        // preserved entry beside an entry replayed from an aging snapshot, so a
-        // per-RECORD gate over-cleared the whole record (the residual documented
-        // at dispatch/integration.ts). A missing per-entry stamp (a migrated
-        // pre-#1186 record) inherits the record's `touchedAt`. Tolerance matches
-        // the Windows host-clock skew rationale above.
-        const mtimeStale = (d) => freshnessFromMtime({
-            mtimeMs,
-            referenceMs: d.observedAt ?? rec.touchedAt,
-        }).verdict === "stale";
-        // #3183: a disposition-tagged row carrying a strict anchor span is gated
-        // on its MARK, not on the file's mtime (retirement rule 1 — see
-        // `WidgetDiagnostic.suppressedRetained`): it survives while the marked
-        // LINE is still somewhere in the file, and is retired when it is not.
-        // The content read is gated on a span-carrying row the mtime gate WOULD
-        // have dropped — the only state where reading can change an outcome — so
-        // the debounced render sweep stays off the disk both for a file with no
-        // marks and for a marked file nothing has touched.
-        const spans = rec.allDiagnostics.some((d) => d.anchorSpan !== undefined && mtimeStale(d))
-            ? readAnchorSpans(rec.filePath)
-            : undefined;
-        const survivors = rec.allDiagnostics.filter((d) => d.anchorSpan !== undefined && spans !== undefined
-            ? spans.has(d.anchorSpan)
-            : !mtimeStale(d));
-        if (survivors.length === rec.allDiagnostics.length) {
-            return { mapKey, action: "keep" }; // nothing stale
-        }
-        if (survivors.length === 0) {
-            return { mapKey, action: "drop" }; // every entry stale → drop record
-        }
-        return { mapKey, action: "prune", survivors };
-    }));
-    let dropped = 0;
-    for (const v of verdicts) {
-        if (v.action === "keep")
-            continue;
-        if (v.action === "drop") {
-            files.delete(v.mapKey);
-            dropped += 1;
-            continue;
-        }
-        // prune: the file changed and shed its stale entries but retains fresher
-        // ones — keep the record, recompute counts/cap from the survivors, and
-        // still count it as a changed file so the agent is told to rescan.
-        const rec = files.get(v.mapKey);
-        if (rec) {
-            rec.allDiagnostics = v.survivors;
-            rec.diagnostics = capStoredDiagnostics(v.survivors);
-            rec.diagnosticCounts = countDiagnostics(v.survivors);
-        }
-        dropped += 1;
-    }
-    if (dropped > 0)
-        requestRenderFn?.();
-    return dropped;
-}
-export async function reconcileStaleWidgetDependencyBlockers(cwd, turnIndex) {
-    let demoted = 0;
-    let truncatedImports = 0;
-    for (const [, rec] of files.entries()) {
-        // #1631 review F4: narrow demotion to LSP-sourced findings. The import-closure
-        // drift check only knows how to reason about IMPORT graphs — an ast-grep
-        // hardcoded-secret or a govulncheck CVE finding doesn't stop being true because
-        // a file this diagnostic's file imports changed; only a language-server verdict
-        // (`tool === "lsp"`, the same predicate `isLspErrorEntry` above uses) is actually
-        // invalidated by that shape of drift. Documented trade: a non-LSP blocking
-        // finding on a file with drifted imports stays fully authoritative until its own
-        // content changes or it is explicitly re-run.
-        if (!rec.allDiagnostics.some((d) => isBlocking(d) && d.tool === "lsp"))
-            continue;
-        let importMtimes = [];
-        try {
-            const result = await collectForwardImportMtimes(cwd, rec.filePath, undefined, turnIndex);
-            importMtimes = result.mtimes;
-            if (result.truncated)
-                truncatedImports += 1;
-        }
-        catch {
-            importMtimes = [];
-        }
-        if (importMtimes.length === 0)
-            continue;
-        let changed = false;
-        for (const d of rec.allDiagnostics) {
-            if (!isBlocking(d) || d.tool !== "lsp")
-                continue;
-            const baseline = d.observedAt ?? rec.touchedAt;
-            // +50ms tolerance (#1631 review F2): a whole-millisecond `Date.now()`
-            // baseline vs. sub-millisecond mtime precision only needs +1ms, but on
-            // Windows a file's mtime can LEAD `Date.now()` by up to ~11.4ms (measured
-            // across 200 writes; #1491/#1498 precedent for the same host-clock skew).
-            // +1ms produced 42 false demotions in 50 runs on Windows; +50ms clears the
-            // measured skew while staying far below the gap between real edits.
-            if (importMtimes.some((im) => freshnessFromMtime({
-                mtimeMs: im.mtimeMs,
-                referenceMs: baseline,
-            }).verdict === "stale")) {
-                d.stale = true;
-                d.staleReason = "dependency-drift";
-                changed = true;
-                demoted += 1;
-            }
-        }
-        if (changed) {
-            rec.diagnosticCounts = countDiagnostics(rec.allDiagnostics);
-            rec.diagnostics = capStoredDiagnostics(rec.allDiagnostics);
-        }
-    }
-    if (demoted > 0)
-        requestRenderFn?.();
-    return { demoted, truncatedImports };
-}
-/**
- * #1790: sweep-population feed for `sweepInlineBlockerFreshness` (the turn-end
- * gate in `clients/blocker-freshness.ts`). That sweep's population was built
- * solely from `RuntimeCoordinator`'s inline-blocker map — a live-dispatch-only
- * store. A workspace-diagnostics CACHE HIT never touches that map; it calls
- * `reconcileScanDiagnostics` straight into `files` here (the cache-serve branch
- * in `tools/lsp-diagnostics.ts`), so a cache-served blocking row could render in
- * the widget while the turn-end sweep's `total` count never saw it (the live
- * 2026-08-20 dogfood: `total:1 kept:1` against five stale cache-served blocking
- * rows on screen).
- *
- * One entry per file with at least one CURRENTLY blocking (non-stale),
- * LSP-sourced diagnostic — the same provenance narrowing
- * `reconcileStaleWidgetDependencyBlockers` already applies, so every row this
- * feeds into the sweep is one its `isLspSourced` gate would keep anyway.
- * `recordedAtMs` is the EARLIEST `observedAt` among them: the conservative
- * baseline, since using the latest could hide drift that predates an earlier
- * diagnostic's own observation.
- */
-export function getWidgetBlockingFilesForSweep() {
-    const out = [];
-    for (const rec of files.values()) {
-        let earliest;
-        for (const d of rec.allDiagnostics) {
-            if (!isBlocking(d) || d.tool !== "lsp")
-                continue;
-            const observedAt = d.observedAt ?? rec.touchedAt;
-            if (earliest === undefined || observedAt < earliest)
-                earliest = observedAt;
-        }
-        if (earliest !== undefined)
-            out.push({ filePath: rec.filePath, recordedAtMs: earliest });
-    }
-    return out;
-}
-/**
- * #1790: demote every currently-blocking, LSP-sourced diagnostic for `filePath`
- * to stale (#1419 demote-not-drop). Called by the turn-end sweep once its own
- * drift check (`detectDrift`/`collectForwardImportMtimes`, shared with the
- * inline-blocker branch) independently confirms drift for a file that reached
- * the sweep's population through `getWidgetBlockingFilesForSweep` rather than an
- * inline blocker. Same store and same write `reconcileStaleWidgetDependencyBlockers`
- * already performs per-diagnostic — this is the single-file entry point the
- * sweep drives once ITS check (not a second implementation of the check) says
- * to. Returns true iff something changed.
- *
- * #1790 review F4 (latent): unlike `reconcileStaleWidgetDependencyBlockers`, which
- * checks drift per-diagnostic against EACH diagnostic's own `observedAt`, this
- * demotes every currently-blocking LSP diagnostic for the file against the ONE
- * file-level baseline `getWidgetBlockingFilesForSweep` computed (the earliest
- * `observedAt` among them) — because the sweep's `detectDrift` contract is
- * file-level, not per-diagnostic. That is safe ONLY as long as every LSP
- * diagnostic on a `FileRecord` shares (or is more conservative than) that
- * earliest stamp, which holds today because every writer that populates
- * `allDiagnostics` (`recordDiagnostics`, `reconcileCascadeNeighborLspErrors`)
- * replaces the file's LSP errors wholesale with one `observedAt` per write. If a
- * future writer ever merges LSP diagnostics with genuinely MIXED per-entry
- * stamps on one record, this function would over-demote entries observed after
- * the earliest one drifted — re-derive per-diagnostic (mirroring
- * `reconcileStaleWidgetDependencyBlockers`) rather than assume the single
- * baseline still holds.
- */
-export function markWidgetFileBlockersStale(filePath, reason) {
-    const rec = files.get(fileMapKey(filePath));
-    if (!rec)
-        return false;
-    let changed = false;
-    for (const d of rec.allDiagnostics) {
-        if (!isBlocking(d) || d.tool !== "lsp")
-            continue;
-        d.stale = true;
-        d.staleReason = reason;
-        changed = true;
-    }
-    if (changed) {
-        rec.diagnosticCounts = countDiagnostics(rec.allDiagnostics);
-        rec.diagnostics = capStoredDiagnostics(rec.allDiagnostics);
-        requestRenderFn?.();
-    }
-    return changed;
-}
-/**
- * #2275: a dependency-drift demotion the footer is still allowed to draw —
- * the one predicate the render marker, the delivery counter and the
- * footer-hide all ask, so "which rows does this cap govern" cannot drift
- * apart between them.
- */
-function isDependencyDriftDemoted(d) {
-    return (d.stale === true &&
-        d.staleReason === "dependency-drift" &&
-        d.footerRetired !== true);
-}
-/**
- * #2275 review F1: take (and clear) the files whose dependency-drift
- * demotions the footer actually RENDERED since the last call — the delivery
- * population `runtime-turn.ts`'s per-turn cap step walks at turn end.
- *
- * Deliberately NOT a store-wide query over everything currently demoted.
- * `renderWidget` serves one record per pass, so a store-wide walk charged a
- * delivery to every demoted file every turn — including files the footer had
- * never shown and one demoted moments earlier by the same turn end, which
- * also made the ledger's "capped after N deliveries" overstate N by one.
- * Draining here is what makes many repaints within one turn count as the one
- * delivery they are, and a turn with no repaint count as none.
- */
-export function drainRenderedDependencyDriftFilePaths() {
-    const out = [...renderedDependencyDriftFiles];
-    renderedDependencyDriftFiles.clear();
-    return out;
-}
-/**
- * #2275: commit one more RENDERED delivery of `filePath`'s
- * dependency-drift-demoted diagnostics. Every qualifying entry on the record
- * advances together — `markWidgetFileBlockersStale` demotes them as one
- * batch per file, so they share one delivery history, mirroring that
- * write's own file-level scope. Returns the new count, or 0 when the file
- * has nothing this cap governs (never demoted, or already footer-retired —
- * a hidden row is delivered no more, so it must stop advancing).
- */
-export function incrementWidgetDependencyDriftDelivery(filePath) {
-    const rec = files.get(fileMapKey(filePath));
-    if (!rec)
-        return 0;
-    let next = 0;
-    for (const d of rec.allDiagnostics) {
-        if (isDependencyDriftDemoted(d)) {
-            next = (d.staleDeliveryCount ?? 0) + 1;
-            d.staleDeliveryCount = next;
-        }
-    }
-    return next;
-}
-/**
- * #2275 review F2: retire every dependency-drift-demoted diagnostic on
- * `filePath` once its delivery cap is reached — by HIDING it from the
- * footer, not by dropping it.
- *
- * `RuntimeCoordinator.retireDemotedDependencyDriftBlocker` can delete its
- * whole `_pendingInlineBlockers` record because that store backs exactly one
- * surface (the turn-end advisory text). This store does not: `allDiagnostics`
- * is also what `getFileDiagnosticSummaries` serves to `lens_diagnostics
- * mode=all` and what `lens_diagnostic_mark` cross-checks to reanchor. Removing
- * the entry there turned an unconfirmed LSP error into a CLEAN read and took
- * it out of the error tally #1631 requires it to stay in. So the cap marks
- * `footerRetired` and leaves everything else — severity, counts, the
- * `stale`/`dependency-drift` demotion — exactly as it was; mode=all carries
- * the "capped, still confirmable" wording in its own note. Returns true iff
- * something changed (idempotent on a second call).
- */
-export function retireWidgetDependencyDriftBlockers(filePath) {
-    const rec = files.get(fileMapKey(filePath));
-    if (!rec)
-        return false;
-    let changed = false;
-    for (const d of rec.allDiagnostics) {
-        if (isDependencyDriftDemoted(d)) {
-            d.footerRetired = true;
-            changed = true;
-        }
-    }
-    if (!changed)
-        return false;
-    // The record's severity tallies are unchanged by design (the finding is
-    // still an error, still non-blocking) — only what the footer draws moves.
-    // Fix-round 3 (#2275 review F1): this delete is live only for the
-    // multi-file drain re-render case — `drainRenderedDependencyDriftFilePaths`
-    // already clears the whole Set before this loop starts, so by the time a
-    // given `wPath` reaches this call its own entry is normally long gone. It
-    // only does something when a render for a DIFFERENT file interleaves
-    // between the drain and this retire (adding `rec.filePath` back to the
-    // Set mid-loop) or a render fires between retirement and the caller's next
-    // drain — either way, defensive: it keeps a just-retired file from being
-    // double-drained (and double-charged) on the next pass.
-    renderedDependencyDriftFiles.delete(rec.filePath);
-    requestRenderFn?.();
-    return true;
-}
-/**
- * Keep the TUI honest (#298 follow-up). `reconcileStaleWidgetFiles` drops
- * widget entries whose file changed on disk after they were last recorded
- * (i.e. diagnostics the agent already fixed) — but it was only ever wired
- * into the `lens_diagnostics` tool, so the widget rendered cached diagnostics
- * verbatim and kept showing fixed errors until `lens_diagnostics` was run by
- * hand. This debounced scheduler fires it from the widget render path (see
- * `mountLensWidget` in index.ts) so stale entries self-correct. The debounce
- * collapses the burst of renders that accompany a save into a single sweep.
- */
-let staleReconcileTimer = null;
-export const STALE_RECONCILE_DEBOUNCE_MS = 1500;
-export function scheduleStaleReconcile() {
-    if (staleReconcileTimer !== null)
-        return;
-    staleReconcileTimer = setTimeout(() => {
-        staleReconcileTimer = null;
-        void reconcileStaleWidgetFiles().catch(() => { });
-    }, STALE_RECONCILE_DEBOUNCE_MS);
-    // Don't keep the process alive solely for this background sweep.
-    staleReconcileTimer?.unref?.();
-}
-/**
- * Return current diagnostics for every file pi-lens has seen this session.
- * Used by lens_diagnostics tool (mode: "all"). Exposes the FULL per-file
- * diagnostic set — decoupled from the widget's display cap — so the agent sees
- * everything, not just the 12 the TUI keeps for rendering.
- *
- * #3183: "everything" means every row a live scan stands behind.
- * `disposition`-tagged rows are the store's own bookkeeping for a mark — the
- * scan that produced this record did not report them — and they are excluded
- * here. That is the same predicate `countDiagnostics` and `isBlocking` have
- * always used, so this makes the projected array agree with the counts beside
- * it; before #3183 the array disagreed, and once retirement rule 1 let a row
- * outlive the file's mtime, `applyCachedDispositions` reached it in the
- * WEAK-fallback branch — which can never match a STRICT `false-positive`
- * anchor — and `mode=all` served a marked finding as a live warning (#3158 AC3).
- * The exclusion lives HERE, at the one projection every cache-only surface and
- * the mode=full merge read, rather than as a second disposition rule inside
- * `summarizeDiagnostics`' tally: that tally still skips `stale` only, and the
- * disposition filter still reaches it pre-applied through the shared
- * `applyCachedDispositions` seam.
- *
- * Deliberately NOT applied to {@link getFileDiagnostics}, the per-file read:
- * `reconcileWidgetDisposition` needs the tagged rows to enforce retirement
- * rule 3, and `lens_diagnostic_mark`'s cross-check re-marks through them.
- */
-export function getFileDiagnosticSummaries() {
-    return [...files.values()].map((rec) => {
-        applyPastEofGate(rec);
-        const live = rec.allDiagnostics.filter((d) => d.disposition === undefined);
-        return {
-            filePath: rec.filePath,
-            blocking: rec.diagnosticCounts.blocking,
-            errors: rec.diagnosticCounts.errors,
-            warnings: rec.diagnosticCounts.warnings,
-            advisories: countAdvisories(live),
-            hasFinalSnapshot: rec.hasFinalDiagnosticsSnapshot,
-            diagnostics: live.map((d) => ({ ...d })),
-        };
-    });
-}
-/**
- * How many {@link WidgetDiagnostic.suppressedRetained} rows this file's record
- * is currently carrying (#3158 round 2 F4).
- *
- * The success-path counterpart of `widget-suppressed-retention-capped`: the
- * ledger only hears about retention when the per-file cap TRUNCATES it, so
- * nothing observed the healthy case. `tools/lsp-diagnostics.ts` folds this into
- * the `lsp_probe_disposition_filter` phase it already emits once per filtered
- * file per scan, so the count is bounded by that record's existing cardinality
- * — never one row per retained finding.
- *
- * Deliberately not {@link getFileDiagnostics}: that applies the past-EOF gate,
- * which can `stat` the file. This is a map lookup plus a count over rows the
- * caller's own scan just filtered, so it adds no I/O to the probe path.
- */
-export function countRetainedSuppressedRows(filePath) {
-    const rec = files.get(fileMapKey(filePath));
-    if (!rec)
-        return 0;
-    let n = 0;
-    for (const d of rec.allDiagnostics)
-        if (d.suppressedRetained)
-            n += 1;
-    return n;
-}
-/**
- * Return the current FULL (uncapped) diagnostic set for a single file, as
- * last recorded by {@link recordDiagnostics} — the same `allDiagnostics`
- * store `getFileDiagnosticSummaries` exposes per-file, without paying for a
- * whole-session snapshot. Used by the #502 `pilens:diagnostics` bus producer
- * (`clients/bus-publish.ts`), which reads this immediately after
- * `recordDiagnostics` writes it so the emitted event reflects the write
- * batch's FINAL diagnostic state (post-format, post-autofix, post-dispatch —
- * see pipeline.ts call order). Returns `undefined` when the file has never
- * been recorded (caller must not confuse "never seen" with "seen and clean";
- * an explicit `[]` from `recordDiagnostics` is a real empty array here).
- *
- * The `files` map key is normalized through `fileMapKey` (#1020), so any path
- * form of the same file — forward-slash, backslash, or a different Windows
- * drive-letter case — resolves to the same record. This read-side fold MUST
- * stay identical to the write-side fold, or a file recorded under one form
- * would silently read as `undefined` under another (e.g. via bus-publish).
- */
-export function getFileDiagnostics(filePath) {
-    const rec = files.get(fileMapKey(filePath));
-    if (!rec)
-        return undefined;
-    applyPastEofGate(rec);
-    return rec.allDiagnostics.map((d) => ({ ...d }));
-}
-/** @internal Test-only helpers. Do not use in production code. */
-export const __testing = {
-    getWidgetStateSnapshot() {
-        return {
-            files: [...files.values()].map((rec) => ({
-                filePath: rec.filePath,
-                storedDiagnostics: rec.diagnostics.length,
-                blocking: rec.diagnosticCounts.blocking,
-                errors: rec.diagnosticCounts.errors,
-                warnings: rec.diagnosticCounts.warnings,
-            })),
-        };
-    },
-};
-export function recordLsp(serverId, root, status, durationMs) {
-    const normalizedRoot = normalizeMapKey(root);
-    const key = `${serverId}@${normalizedRoot}`;
-    const mapped = status === "spawn_start"
-        ? "spawning"
-        : status === "spawn_success"
-            ? "ready"
-            : "failed";
-    lspServers.set(key, { serverId, root, status: mapped, durationMs });
-    requestRender();
-}
-// ── Render ────────────────────────────────────────────────────────────────────
-const HORIZONTAL_MIN_WIDTH = 70;
-export function renderWidget(width, theme) {
-    const dim = (s) => theme.fg("dim", s);
-    const red = (s) => theme.fg("error", s);
-    const yellow = (s) => theme.fg("warning", s);
-    const green = (s) => theme.fg("success", s);
-    const cyan = (s) => theme.fg("accent", s);
-    const w = Math.max(1, width || 80);
-    const useHorizontal = w >= HORIZONTAL_MIN_WIDTH;
-    if (files.size === 0 && lspServers.size === 0)
-        return [];
-    const lines = [];
-    // Header — counts from deduplicated files only
-    const deduped = dedupeByBasename([...files.values()]);
-    const recencySorted = deduped.filter(shouldRenderFile).slice(0, 5);
-    const langStr = sessionLanguages.slice(0, 6).join(" ");
-    const totalBlocking = countBlockingIn(deduped);
-    const totalErrors = countTotalIn("error", deduped);
-    const totalWarnings = countTotalIn("warning", deduped);
-    const totalSuppressed = countSuppressedIn(deduped);
-    const hasPendingAnalysis = deduped.some(isPendingAnalysis);
-    const errorChunk = totalErrors > 0
-        ? (totalBlocking > 0 ? red : yellow)(`●${totalErrors}E`)
-        : "";
-    const warningChunk = totalWarnings > 0 ? yellow(`!${totalWarnings}W`) : "";
-    const summary = errorChunk
-        ? errorChunk + (warningChunk ? " " + warningChunk : "")
-        : warningChunk
-            ? warningChunk
-            : files.size > 0 && !hasPendingAnalysis
-                ? green("✓ clean")
-                : "";
-    // LSP spawning — folded into the header in horizontal mode, tail line otherwise
-    const spawning = [...lspServers.values()].filter((s) => s.status === "spawning");
-    const lspChip = useHorizontal && spawning.length > 0 ? "  " + dim("LSP↑") : "";
-    const header = ` ${cyan("pi-lens")}${langStr ? "  " + dim(langStr) : ""}${lspChip}${summary ? "  " + summary : ""}`;
-    lines.push(fitLine(header, w));
-    if (totalSuppressed > 0) {
-        lines.push(fitLine(` ${dim(`suppressed: ${totalSuppressed}`)}`, w));
-    }
-    // File list — display order varies by mode
-    if (useHorizontal) {
-        const displayOrder = sortByTierThenRecency(recencySorted);
-        const rowLine = packHorizontalRow(displayOrder, w, theme);
-        if (rowLine.length > 0)
-            lines.push(rowLine);
-    }
-    else {
-        for (const rec of recencySorted) {
-            lines.push(fitLine(formatFileRowVertical(rec, theme), w));
-        }
-    }
-    // Diagnostics — blocking (or dependency-drift-demoted) only, from the most
-    // recently touched file that has them. Vertical mode keeps the divider/filename
-    // context; horizontal already shows the filename on the packed row above, so
-    // we drop the extra header noise there.
-    //
-    // #1631 review F10: `isBlocking` answers false for a demoted (`stale`) finding
-    // by design (#1419 demote-not-drop) — but this render loop used that same
-    // predicate to pick which findings to SHOW, so a demoted finding vanished from
-    // the footer entirely instead of demoting visibly. `isBlockingOrDemoted` keeps
-    // it in the list; the per-entry render below distinguishes the two with a
-    // dimmed marker instead of the red dot.
-    //
-    // #2275: a demotion that reached its delivery cap is hidden from THIS
-    // surface only (`footerRetired`) — it stays in `allDiagnostics` for
-    // mode=all and lens_diagnostic_mark, which is where the agent is told it
-    // was capped and can still be confirmed by a re-run.
-    const isBlockingOrDemoted = (d) => (isBlocking(d) || d.stale === true) && d.footerRetired !== true;
-    const withBlocking = recencySorted.filter((r) => r.diagnostics.some(isBlockingOrDemoted));
-    if (withBlocking.length > 0) {
-        const rec = withBlocking[0];
-        // #1641: gate only the ONE record whose line numbers are about to be
-        // rendered, not the whole (potentially session-long) `deduped` list the
-        // header counts read — one memoized stat per redraw, never a per-file
-        // scan of every file pi-lens has touched this session. A past-EOF
-        // citation demoted here can leave the header's aggregate blocking count
-        // one turn stale; `reconcileStaleWidgetFiles`'s existing debounced sweep
-        // (scheduleStaleReconcile) already re-derives that aggregate on its own
-        // cadence, so this is a bounded, self-correcting gap, not a silent one.
-        applyPastEofGate(rec);
-        if (!useHorizontal) {
-            lines.push(fitLine(dim("─".repeat(Math.min(w, 60))), w));
-            lines.push(fitLine(` ${dim(path.basename(rec.filePath))}`, w));
-        }
-        const blockers = rec.diagnostics.filter(isBlockingOrDemoted).slice(0, 5);
-        // #2275 review F1: THIS is the point where a demoted row is delivered to
-        // a human — after the `withBlocking[0]` pick and the top-five slice, not
-        // before them. Mark the file so the next turn end can charge exactly one
-        // delivery against its cap. File-level, matching
-        // `markWidgetFileBlockersStale`'s own file-level demotion batch (all of a
-        // file's drift demotions share one delivery history); a file holding more
-        // than five demoted rows delivers the visible ones and counts the batch,
-        // which is the conservative direction — it retires no LATER than the rows
-        // the agent actually read.
-        if (blockers.some(isDependencyDriftDemoted)) {
-            renderedDependencyDriftFiles.add(rec.filePath);
-        }
-        for (const d of blockers) {
-            // A past-EOF demotion's coordinate is untrustworthy — render the
-            // marker instead of the line (#1641); drift demotions keep theirs.
-            const pastEof = d.stale && (d.staleReason ?? "past-eof") === "past-eof";
-            // No link for a past-EOF demotion: the anchor would carry the same
-            // untrustworthy line the marker exists to replace.
-            const loc = pastEof
-                ? PAST_EOF_STALE_MARKER
-                : d.line != null
-                    ? osc8(d.uri ?? "", `L${d.line}`)
-                    : "";
-            const rule = d.rule ? dim(` ${d.rule}`) : "";
-            const staleTag = d.stale ? dim(` ${STALE_LINE_MARKER}`) : "";
-            const prefix = `   ${d.stale ? dim("○") : red("●")} ${loc}${rule}  `;
-            const msgWidth = Math.max(1, w - visibleWidth(prefix) - visibleWidth(staleTag));
-            const msg = fitLine(d.message, msgWidth, "…");
-            lines.push(fitLine(`${prefix}${msg}${staleTag}`, w));
-        }
-    }
-    // LSP status tail — only in vertical mode; horizontal folds into header
-    if (!useHorizontal && spawning.length > 0) {
-        const ids = spawning.map((s) => s.serverId).join(" ");
-        lines.push(fitLine(` ${dim(`LSP spawning: ${ids}`)}`, w));
-    }
-    return lines;
-}
-function classifyFileTier(rec) {
-    if (rec.diagnosticCounts.blocking > 0)
-        return "blocking";
-    if (rec.diagnosticCounts.errors > 0 || rec.diagnosticCounts.warnings > 0) {
-        return "warning";
-    }
-    return "clean";
-}
-function sortByTierThenRecency(recs) {
-    const order = {
-        blocking: 0,
-        warning: 1,
-        clean: 2,
-    };
-    return [...recs].sort((a, b) => {
-        const ta = order[classifyFileTier(a)];
-        const tb = order[classifyFileTier(b)];
-        if (ta !== tb)
-            return ta - tb;
-        return b.touchedAt - a.touchedAt;
-    });
-}
-function formatFileRowVertical(rec, theme) {
-    const dim = (s) => theme.fg("dim", s);
-    const red = (s) => theme.fg("error", s);
-    const yellow = (s) => theme.fg("warning", s);
-    const green = (s) => theme.fg("success", s);
-    const base = path.basename(rec.filePath);
-    const blocking = rec.diagnosticCounts.blocking;
-    const errors = rec.diagnosticCounts.errors;
-    const warnings = rec.diagnosticCounts.warnings;
-    const formatterFailed = hasFailedFormatter(rec);
-    // Diagnostic severity outranks formatter failure (#1348 review): a file
-    // with blocking diagnostics shows the blocking dot even if a format also
-    // failed -- same precedence as the horizontal renderer.
-    const dot = blocking > 0
-        ? red("●")
-        : formatterFailed
-            ? red("x")
-            : warnings > 0 || errors > 0
-                ? yellow("!")
-                : green("✓");
-    const runnerNames = [...rec.runners.entries()]
-        .filter(([, r]) => r.status !== "skipped")
-        .map(([id]) => id)
-        .join(" ");
-    const counts = errors > 0
-        ? " " +
-            (blocking > 0 ? red : yellow)(`${errors}E`) +
-            (warnings > 0 ? " " + yellow(`${warnings}W`) : "")
-        : warnings > 0
-            ? " " + yellow(`${warnings}W`)
-            : " " + dim("clean");
-    const changedFormatters = [...rec.formatters.entries()]
-        .filter(([, f]) => f.changed && f.success)
-        .map(([name]) => name);
-    const failedFormatters = [...rec.formatters.entries()]
-        // An `unavailable` tool is not a code failure (#2413): it never earns a red
-        // `fmt-failed` marker. Its result already carries success:true, so this
-        // guard is a belt that keeps the intent explicit at the render seam.
-        .filter(([, f]) => !f.success && f.outcome !== "unavailable")
-        .map(([name]) => name);
-    const formatMark = (failedFormatters.length > 0
-        ? red(` fmt-failed:${failedFormatters.join(",")}`)
-        : "") +
-        (changedFormatters.length > 0
-            ? dim(` fmt:${changedFormatters.join(",")}`)
-            : "");
-    return ` ${dot} ${base}  ${dim(runnerNames)}${formatMark}${counts}`;
-}
-function packHorizontalRow(recs, totalWidth, theme) {
-    if (recs.length === 0)
-        return "";
-    const dim = (s) => theme.fg("dim", s);
-    const indent = "   ";
-    const sep = "  ";
-    // Reserve worst-case overflow space upfront so the marker always fits.
-    // " +NN" — 4 visible chars covers up to two-digit overflow.
-    const overflowReserve = 4;
-    let used = visibleWidth(indent);
-    const parts = [indent];
-    const addedTokenWidths = [];
-    let droppedAt = -1;
-    for (let i = 0; i < recs.length; i++) {
-        const sepWidth = parts.length > 1 ? visibleWidth(sep) : 0;
-        const willOverflow = i < recs.length - 1;
-        const reserve = willOverflow ? overflowReserve : 0;
-        const remaining = totalWidth - used - sepWidth - reserve;
-        if (remaining < 4) {
-            droppedAt = i;
-            break;
-        }
-        const token = formatFileTokenHorizontal(recs[i], remaining, theme);
-        const tokenWidth = visibleWidth(token);
-        if (token.length === 0 || used + sepWidth + tokenWidth > totalWidth) {
-            droppedAt = i;
-            break;
-        }
-        if (sepWidth > 0) {
-            parts.push(sep);
-            used += sepWidth;
-        }
-        parts.push(token);
-        used += tokenWidth;
-        addedTokenWidths.push(tokenWidth + sepWidth);
-    }
-    if (droppedAt >= 0) {
-        let dropped = recs.length - droppedAt;
-        let overflow = " " + dim(`+${dropped}`);
-        // If reservation was insufficient (e.g. last token grew because no
-        // reserve was applied), shed accepted tokens until overflow fits.
-        while (used + visibleWidth(overflow) > totalWidth &&
-            addedTokenWidths.length > 0) {
-            const lastWidth = addedTokenWidths.pop();
-            used -= lastWidth;
-            parts.pop(); // token
-            if (parts.length > 1)
-                parts.pop(); // preceding separator
-            dropped++;
-            overflow = " " + dim(`+${dropped}`);
-        }
-        if (used + visibleWidth(overflow) <= totalWidth) {
-            parts.push(overflow);
-        }
-    }
-    return fitLine(parts.join(""), totalWidth);
-}
-function formatFileTokenHorizontal(rec, remainingWidth, theme) {
-    const dim = (s) => theme.fg("dim", s);
-    const red = (s) => theme.fg("error", s);
-    const yellow = (s) => theme.fg("warning", s);
-    const blocking = rec.diagnosticCounts.blocking;
-    const errors = rec.diagnosticCounts.errors;
-    const warnings = rec.diagnosticCounts.warnings;
-    const formatterChanged = hasChangedFormatter(rec);
-    const formatterFailed = hasFailedFormatter(rec);
-    let dotChar;
-    if (blocking > 0)
-        dotChar = red("●");
-    else if (errors > 0 || warnings > 0)
-        dotChar = yellow("!");
-    else if (formatterChanged)
-        dotChar = dim("✎");
-    else
-        dotChar = dim("·");
-    if (formatterFailed && blocking === 0 && errors === 0 && warnings === 0) {
-        dotChar = red("x");
-    }
-    let countsStyled = "";
-    if (errors > 0 && warnings > 0) {
-        const eColor = blocking > 0 ? red : yellow;
-        countsStyled = " " + eColor(`${errors}E`) + yellow(`${warnings}W`);
-    }
-    else if (errors > 0) {
-        const eColor = blocking > 0 ? red : yellow;
-        countsStyled = " " + eColor(`${errors}E`);
-    }
-    else if (warnings > 0) {
-        countsStyled = " " + yellow(`${warnings}W`);
-    }
-    const fullBasename = path.basename(rec.filePath);
-    const fixedWidth = visibleWidth(dotChar) + 1 + visibleWidth(countsStyled);
-    const basenameBudget = remainingWidth - fixedWidth;
-    if (basenameBudget < 3)
-        return "";
-    const truncated = truncateBasename(fullBasename, basenameBudget);
-    const linked = osc8(pathToFileURL(rec.filePath).href, truncated);
-    return `${dotChar} ${linked}${countsStyled}`;
-}
-function truncateBasename(name, maxWidth) {
-    if (visibleWidth(name) <= maxWidth)
-        return name;
-    if (maxWidth < 2)
-        return "…";
-    const ext = path.extname(name);
-    const stem = name.slice(0, name.length - ext.length);
-    const keep = maxWidth - ext.length - 1;
-    if (keep < 1) {
-        // Extension alone wouldn't fit; truncate the whole name.
-        return name.slice(0, maxWidth - 1) + "…";
-    }
-    return stem.slice(0, keep) + "…" + ext;
-}
-// ── Helpers ──────────────────────────────────────────────────────────────────
-function getOrCreate(filePath, key = fileMapKey(filePath)) {
-    maybePruneInactiveFileRecords();
-    // Look up by the normalized key so mixed path forms of the same file share
-    // ONE record (#1020); keep the caller's verbatim path as the display path.
-    return (files.get(key) ?? {
-        filePath,
-        runners: new Map(),
-        formatters: new Map(),
-        diagnostics: [],
-        allDiagnostics: [],
-        diagnosticCounts: { blocking: 0, errors: 0, warnings: 0 },
-        hasFinalDiagnosticsSnapshot: false,
-        touchedAt: Date.now(),
-    });
-}
-function hasChangedFormatter(rec) {
-    return [...rec.formatters.values()].some((f) => f.changed && f.success);
-}
-function hasFailedFormatter(rec) {
-    // `unavailable` (#2413) is success:true and must never count as a failure —
-    // the explicit guard keeps that true even if a future path mislabels it.
-    return [...rec.formatters.values()].some((f) => !f.success && f.outcome !== "unavailable");
-}
-function shouldRenderFile(rec) {
-    return (rec.hasFinalDiagnosticsSnapshot ||
-        hasChangedFormatter(rec) ||
-        hasFailedFormatter(rec));
-}
-function hasLiveDiagnostic(rec) {
-    return rec.hasFinalDiagnosticsSnapshot && rec.diagnostics.length > 0;
-}
-function isPendingAnalysis(rec) {
-    return rec.runners.size > 0 && !rec.hasFinalDiagnosticsSnapshot;
-}
-function capStoredDiagnostics(diagnostics) {
-    if (diagnostics.length <= MAX_STORED_DIAGNOSTICS_PER_FILE)
-        return diagnostics;
-    const blockers = diagnostics.filter(isBlocking);
-    if (blockers.length >= MAX_STORED_DIAGNOSTICS_PER_FILE) {
-        return blockers.slice(0, MAX_STORED_DIAGNOSTICS_PER_FILE);
-    }
-    const rest = diagnostics.filter((d) => !isBlocking(d));
-    return [
-        ...blockers,
-        ...rest.slice(0, MAX_STORED_DIAGNOSTICS_PER_FILE - blockers.length),
-    ];
-}
-function countTotalIn(severity, recs) {
-    let n = 0;
-    for (const rec of recs) {
-        if (severity === "error")
-            n += rec.diagnosticCounts.errors;
-        else
-            n += rec.diagnosticCounts.warnings;
-    }
-    return n;
-}
-function countSuppressedIn(recs) {
-    let n = 0;
-    for (const rec of recs) {
-        for (const diagnostic of rec.allDiagnostics) {
-            if (diagnostic.disposition)
-                n++;
-        }
-    }
-    return n;
-}
-function countBlockingIn(recs) {
-    let n = 0;
-    for (const rec of recs)
-        n += rec.diagnosticCounts.blocking;
-    return n;
-}
-function requestRender() {
-    requestRenderFn?.();
-}
-function osc8(uri, label) {
-    if (!uri)
-        return label;
-    return `\x1b]8;;${uri}\x1b\\${label}\x1b]8;;\x1b\\`;
-}
-// Dual-signature truncateToWidth handling lives in tui-fit.ts (shared with the
-// turn-summary message renderer, which learned the hard way that pi-tui crashes
-// the host on over-width lines — #513).
-function dedupeByBasename(recs) {
-    const seen = new Map();
-    for (const r of [...recs].sort((a, b) => a.touchedAt - b.touchedAt)) {
-        seen.set(path.basename(r.filePath), r);
-    }
-    return [...seen.values()].sort((a, b) => b.touchedAt - a.touchedAt);
-}