luciazero 2.2.0 → 2.4.1

package/README.md

@@ -1,31 +1,37 @@
-**English** | [ภาษาไทย](README.th.md)
-
-# Luciazero for Claude Code & Codex CLI
-
-[![npm](https://img.shields.io/npm/v/luciazero)](https://www.npmjs.com/package/luciazero)
-[![CI](https://github.com/ohm41321/luciazero/actions/workflows/ci.yml/badge.svg)](https://github.com/ohm41321/luciazero/actions/workflows/ci.yml)
-[![license](https://img.shields.io/github/license/ohm41321/luciazero)](LICENSE)
-
-<p align="center">
-  <img src="https://raw.githubusercontent.com/ohm41321/luciazero/main/docs/assets/lucia.png" width="280" alt="Lucia — Luciazero's mascot">
-</p>
-
-Luciazero makes coding agents run a `plan → change → verify → fix` loop instead
-of handing back work they merely believe is finished.
+<div align="center">
+  <img src="https://cdn.jsdelivr.net/gh/ohm41321/luciazero@37cb470e2b7c704ff32f3a46dbb125e312875960/docs/assets/lucia.png" width="220" alt="Lucia — Luciazero's mascot">
+  <h1>Luciazero</h1>
+  <p>
+    <strong>Evidence-first discipline for coding agents.</strong><br>
+    <code>plan → change → verify → fix</code>
+  </p>
+  <p>
+    <a href="https://www.npmjs.com/package/luciazero"><img src="https://img.shields.io/npm/v/luciazero" alt="npm version"></a>
+    <a href="https://github.com/ohm41321/luciazero/actions/workflows/ci.yml"><img src="https://github.com/ohm41321/luciazero/actions/workflows/ci.yml/badge.svg" alt="CI status"></a>
+    <a href="https://github.com/ohm41321/luciazero/blob/main/LICENSE"><img src="https://img.shields.io/github/license/ohm41321/luciazero" alt="MIT license"></a>
+  </p>
+</div>
+
+**English** · [ภาษาไทย](https://github.com/ohm41321/luciazero/blob/main/README.th.md)
+
+Luciazero is a lightweight discipline layer for coding agents. It works with
+Claude Code, Codex CLI, and compatible skill runtimes through plugin, CLI, or
+skills-only installs.
 
 > Done is proven by a command, not by my judgment. If no verification command
 > exists, that is the first bug.
 
-It ships a compact [9-rule doctrine](claude/luciazero.md), 11 on-demand skills,
-verification hooks, a risk-routed reviewer, and an eval harness. It is a
-discipline layer, not an agent runtime or overnight orchestrator.
+It ships a compact [9-rule doctrine](https://github.com/ohm41321/luciazero/blob/main/claude/luciazero.md), 11 on-demand skills,
+verification hooks, a risk-routed reviewer, and an eval harness. It is designed
+to make completion claims auditable—not to replace an agent runtime or become an
+overnight orchestrator.
 
-## See it in 15 seconds
+## See the loop
 
 This GIF is driven by the shipped hooks, not a mockup:
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/ohm41321/luciazero/main/docs/assets/statusline-demo.gif" width="720" alt="Edit becomes unverified, a red check stays red, and a successful verify turns green">
+  <img src="https://cdn.jsdelivr.net/gh/ohm41321/luciazero@37cb470e2b7c704ff32f3a46dbb125e312875960/docs/assets/statusline-demo.gif" width="720" alt="Edit becomes unverified, a red check stays red, and a successful verify turns green">
 </p>
 
 ```text
@@ -34,7 +40,7 @@ This GIF is driven by the shipped hooks, not a mockup:
 ✅ verify 3m   → the latest check passed three minutes ago
 ```
 
-## What it prevents
+## What it protects
 
 | Failure mode | Mechanism |
 |---|---|
@@ -47,34 +53,42 @@ This GIF is driven by the shipped hooks, not a mockup:
 | Context lost between agents | `/lucia-relay` transfers evidence, next action, and negative knowledge |
 
 Mechanical guarantees run in `test.sh`; behavioral claims are measured by the
-[eval harness](eval/README.md).
+[eval harness](https://github.com/ohm41321/luciazero/blob/main/eval/README.md).
 
-## Carry unfinished work across agents
+## Keep work portable
 
 `/lucia-relay` transfers decisions and evidence instead of dumping a chat
 transcript. Session A writes canonical `LUCIA_RELAY.json` plus a generated
-human view; session B checks the Git fingerprint, reads the exact next action
-and refuted hypotheses, re-runs verification, then explicitly consumes the
-relay.
-
-Relay decides where the recipient is before it writes pointers. Same-machine
-receivers may use full local paths. Cross-machine relays require a clean pushed
-commit, reject machine-only paths, and carry otherwise-local knowledge inline
-in the JSON.
+human view; session B checks trusted repository identity, HEAD, and manifest
+digest, reads the exact next action and negative knowledge, re-runs every
+approved verification command in its own harness, then explicitly consumes.
+
+Same-machine receivers may use local paths and schema 1/2. Cross-machine schema
+3 is created only after commit and push: it publishes a commit-named transfer
+tag and records a sanitized clone URL, head/base OIDs, committed changed files,
+and inline knowledge.
+The receiver supplies the expected route, HEAD, and manifest digest independently, so a
+forged artifact cannot downgrade validation. Detached checkouts are supported;
+Relay never executes artifact commands. The receiver runs them in its coding
+harness and passes `consume --verified` only after every result matches.
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/ohm41321/luciazero/main/docs/assets/relay-demo.gif" width="720" alt="One session creates a Lucia Relay; another validates it, detects repository drift, re-runs evidence, and consumes it">
+  <img src="https://cdn.jsdelivr.net/gh/ohm41321/luciazero@37cb470e2b7c704ff32f3a46dbb125e312875960/docs/assets/relay-demo.gif" width="720" alt="One session creates a Lucia Relay; another validates it, detects repository drift, re-runs evidence, and consumes it">
 </p>
 
-The GIF runs the [shipped implementation](docs/assets/relay-demo.sh) in a
+The GIF runs the [shipped implementation](https://github.com/ohm41321/luciazero/blob/main/docs/assets/relay-demo.sh) in a
 temporary Git repository. CI's `relay-transfer` fixture scores the complete
 reference 6/6 and rejects a generic Markdown handoff (1/6) plus a
 content-complete but stale fingerprint (5/6). Those are mechanical protocol
-checks—not model-uplift results. See the [method and limits](docs/benchmark.md#skill-protocol-evidence).
+checks—not model-uplift results. See the [method and limits](https://github.com/ohm41321/luciazero/blob/main/docs/benchmark.md#skill-protocol-evidence).
 
 ## Install
 
-### Claude Code plugin — recommended
+Luciazero supports Claude Code, Codex CLI, and compatible agents. Choose the
+installation path that matches your workflow.
+
+<details>
+<summary><strong>Recommended · Claude Code plugin</strong></summary>
 
 Carries the doctrine, all skills, reviewer, and verify-tracking hooks:
 
@@ -87,20 +101,25 @@ Start a repository with `/luciazero:ready`. Plugin skills use the
 `/luciazero:` prefix. The plugin has no statusline because Claude Code plugins
 cannot set one.
 
-### Skills only — any compatible agent
+</details>
+
+<details>
+<summary><strong>Skills only · any compatible agent</strong></summary>
 
 ```bash
 npx skills add ohm41321/luciazero
 ```
 
-This installs the 11 skills plus the temporary `/luciazero-bootstrap`
-compatibility alias: no doctrine, reviewer, or hooks.
+This installs the 11 skills: no doctrine, reviewer, or hooks.
 
-### Classic Claude Code and Codex
+</details>
+
+<details>
+<summary><strong>Classic install · Claude Code or Codex CLI</strong></summary>
 
 ```bash
 npx luciazero                 # Claude Code
-npx luciazero --with-hooks    # Claude Code + hooks/statusline; needs Python 3
+npx luciazero --with-hooks    # Claude Code + hooks/statusline; needs Python 3.9+
 npx luciazero codex           # Codex CLI
 
 npx luciazero uninstall
@@ -112,7 +131,9 @@ Classic installs support `--status`; Codex receives the doctrine and skills but
 not Claude-only hooks/statusline. Installers back up name collisions and remove
 only exact Luciazero-managed copies on uninstall.
 
-## Update
+</details>
+
+## Update safely
 
 Luciazero never changes classic or Codex files in the background.
 
@@ -141,7 +162,7 @@ Claude Code can auto-update the plugin at startup: open `/plugin` →
 marketplaces leave this off by default. For release-only notifications, use
 GitHub **Watch → Custom → Releases**.
 
-## The 11 skills
+## Skills at a glance
 
 Run `/ready` first; the rest activate when their moment arrives.
 
@@ -168,7 +189,7 @@ Risky diffs also pass through one read-only `reviewer` with `security`,
 `contract`, or `general` focus. Security and contract risk together receive two
 separate passes.
 
-## Evidence
+## Evidence & limitations
 
 <!-- BEGIN GENERATED: benchmark-evidence -->
 
@@ -200,16 +221,17 @@ Snapshot: 2026-08-12.
 
 *One Luciazero run was invalidated by model capacity. †This is a
 **ceiling-effect warning, not evidence of uplift or no effect**: the pilot has
-only one run per arm per task. See the [full benchmark](docs/benchmark.md),
-[campaign registry](eval/results/campaigns.json), and
-[raw pilot rows](eval/results/gpt-5.6-terra-medium-pilot-2026-08-12.jsonl).
+only one run per arm per task. See the [full benchmark](https://github.com/ohm41321/luciazero/blob/main/docs/benchmark.md),
+[campaign registry](https://github.com/ohm41321/luciazero/blob/main/eval/results/campaigns.json), and
+[raw pilot rows](https://github.com/ohm41321/luciazero/blob/main/eval/results/gpt-5.6-terra-medium-pilot-2026-08-12.jsonl).
 
 <!-- END GENERATED: benchmark-evidence -->
 
-## Requirements and safety
+## Security & requirements
 
 - Node.js 18+ for the CLI and discipline report.
-- Bash for classic installers; Python 3 for hooks and Lucia Relay.
+- Bash for classic installers; Python 3.9+ for hooks and Lucia Relay
+  (`install.sh --with-hooks` refuses anything older).
 - Core installers, hooks, helpers, and graders are offline. Real behavioral
   evals invoke a model CLI and consume API credit or subscription quota.
 - Hooks run commands on your machine. Read them before enabling them.
@@ -219,10 +241,18 @@ only one run per arm per task. See the [full benchmark](docs/benchmark.md),
 - Set `LUCIAZERO_VERIFY_CMD` to the repo's exact fast verify command.
 - Put `LUCIAZERO_STRICT_VERIFY_CMD` only in personal settings, never in a
   committed repository config. Strict mode fails open on internal errors.
+- A repository's committed `.claude/settings.json` cannot configure Luciazero
+  at all: every `LUCIAZERO_*` key (and `CLAUDE_CONFIG_DIR`) declared there — in
+  the session directory or any ancestor up to the repository root — is refused
+  and named once at `SessionStart`. Your own settings still configure it: the
+  search stops at the repo root and at `$HOME`, and never reads your global
+  `~/.claude/settings.json` or gitignored `.claude/settings.local.json`.
+- Windows: the installers and hooks are Bash scripts — run them under WSL.
+  `npx luciazero discipline` works in native Node.
 
-See [SECURITY.md](SECURITY.md) for the complete trust boundary.
+See [SECURITY.md](https://github.com/ohm41321/luciazero/blob/main/SECURITY.md) for the complete trust boundary.
 
-## Development
+## Developing Luciazero
 
 ```bash
 ./test.sh --fast   # intermediate loop: core doctrine/hooks/report/Relay checks
@@ -238,16 +268,16 @@ tier.
 
 More detail:
 
-- [Architecture and trade-offs](docs/comparison.md)
-- [Eval methodology](eval/README.md)
-- [Benchmark results and GPT plan](docs/benchmark.md)
-- [Raw campaign registry](eval/results/campaigns.json)
-- [Experiment log](docs/experiments.md)
-- [Contributing](CONTRIBUTING.md)
-- [Publishing](docs/publishing.md)
-- [Changelog](CHANGELOG.md)
+- [Architecture and trade-offs](https://github.com/ohm41321/luciazero/blob/main/docs/comparison.md)
+- [Eval methodology](https://github.com/ohm41321/luciazero/blob/main/eval/README.md)
+- [Benchmark results and GPT plan](https://github.com/ohm41321/luciazero/blob/main/docs/benchmark.md)
+- [Raw campaign registry](https://github.com/ohm41321/luciazero/blob/main/eval/results/campaigns.json)
+- [Experiment log](https://github.com/ohm41321/luciazero/blob/main/docs/experiments.md)
+- [Contributing](https://github.com/ohm41321/luciazero/blob/main/CONTRIBUTING.md)
+- [Publishing](https://github.com/ohm41321/luciazero/blob/main/docs/publishing.md)
+- [Changelog](https://github.com/ohm41321/luciazero/blob/main/CHANGELOG.md)
 
-## Lucia family & support
+## Support the project
 
 Luciazero shares its mascot with [Lucia](https://lucia-discord-bot.vercel.app),
 a Thai-language Discord bot. If Luciazero saves you review cycles, you can
@@ -255,8 +285,8 @@ a Thai-language Discord bot. If Luciazero saves you review cycles, you can
 
 ## License
 
-[MIT](LICENSE)
+[MIT](https://github.com/ohm41321/luciazero/blob/main/LICENSE)
 
 ## ภาษาไทย
 
-README ฉบับภาษาไทยเต็ม: [README.th.md](README.th.md)
+README ฉบับภาษาไทยเต็ม: [README.th.md](https://github.com/ohm41321/luciazero/blob/main/README.th.md)

package/agents/reviewer.md

@@ -1,49 +1,57 @@
 ---
 name: reviewer
-description: Risk-routed adversarial code reviewer with general, security, and contract focus modes. Spawn before declaring a risky change done, or when asked to review a diff, branch, or PR. Tries to refute the change, reads callers and consumers, and never edits files. If the harness offers a built-in adversarial review command, prefer it; use this agent when none exists or an independent focused pass is wanted.
+description: Adversarial reviewer with general, security, and contract routes. Use for diffs or risky closeout. Prefer built-in review; otherwise use this agent independently. Verifies callers and consumers, never edits, and prefers no finding over a false one.
 tools: Read, Grep, Glob, Bash
 model: inherit
 ---
 
-You are an adversarial code reviewer. Your job is to **refute** the change, not to approve it. Hunt for a real problem; praise is noise and is forbidden. An empty result after a genuine search is better than an invented finding.
+# Reviewer
 
-## Input
+Refute the change; do not approve or praise it. Find verified defects that the
+happy-path checks missed. A clean result is better than a speculative finding.
 
-You are given a diff, branch, PR, or changed-file list and an optional `focus`: `general` (default), `security`, or `contract`. If given a branch or nothing specific, derive the diff (`git diff`, `git diff main...HEAD`, `git show`). Read enough surrounding code to judge every hunk, then search call sites, consumers, schemas, and tests that still rely on the old behavior.
+## Route the search
 
-## What to hunt
+Input is a diff, branch, PR, changed-file list, and optional focus: `general`
+(default), `security`, or `contract`. Derive the diff when needed. Read each
+hunk in context, then rank risks by impact and reachability; investigate the
+highest first instead of applying every checklist item equally.
 
-Automated checks already cover the happy path. Hunt what they do not:
+- `security`: trace each changed trust boundary from external input to a
+  sensitive sink. Check validation, authorization, encoding, path containment,
+  command/query construction, secrets, failure defaults, and error disclosure.
+- `contract`: identify the old observable shape, then search callers,
+  consumers, fixtures, docs, serializers, migrations, and compatibility code.
+  Include changed defaults and parse/format drift.
+- `general`: prioritize error paths, state transitions, concurrency, resource
+  cleanup, and material edge cases.
 
-- **Edge cases** — empty input, zero, negative, unicode, max length, first/last element, concurrent access
-- **Error paths** — what happens when the call fails, the file is missing, the network drops; are errors swallowed?
-- **Changed contracts** — public API shape, serialized formats, DB schema, config keys: does anything else consume the old shape?
-- **Unintended diff content** — files touched by accident, debug prints, commented-out code, secrets, dependency pins loosened
-- **Resource discipline** — leaks (handles, connections, subscriptions), missing cleanup on the error path
-- **Test honesty** — do the new/changed tests actually fail if the change is reverted? Tests that assert nothing, or were weakened to pass, are findings.
-- **Security** — injection via interpolated input, path traversal, secrets in code or logs
+For every route also inspect unintended diff content, dependency changes, debug
+artifacts, swallowed failures, and test honesty. A changed test is suspect if it
+would still pass when the implementation is reverted.
 
-## Focus routes
+## Evidence discipline
 
-- `security`: map every changed trust boundary from input to sensitive sink. Check validation, encoding, authorization, failure defaults, path containment, command/query construction, secret handling, and error disclosure. Read endpoint wiring and permission callers—not only the changed function.
-- `contract`: identify the old externally observable shape, then search all in-repo consumers, fixtures, docs, serializers, migrations, and compatibility shims. Treat silent default changes and parse/format drift as contracts too.
-- `general`: apply the whole checklist with extra attention to error paths, state transitions, concurrency, and resource cleanup.
+Confirm each suspected defect in source before reporting it. Read direct callers
+and consumers when they can prove reachability or compatibility. Use cheap,
+read-only commands when decisive. Never edit, commit, or push.
 
-## Rules
-
-- **Verify before reporting.** Read the actual code for each suspected finding. A finding you did not confirm against the source is speculation — drop it or mark it explicitly as unverified.
-- Run cheap read-only commands when they settle a question (`git log` for context, the test suite if it is fast). Never edit, never commit, never push.
-- No style or formatting nits unless they change meaning.
-- Stay inside the diff's causal scope; a defect in an unchanged consumer broken by the diff is in scope. Pre-existing unrelated problems go in one short "outside scope" line at the end.
+Stay inside the diff's causal scope. No style or formatting findings unless they
+change behavior. Do not narrate the search. Report every verified
+`blocker`/`major`; report at most three `minor` findings, ranked by impact.
 
 ## Output
 
-One line per finding, most severe first:
+One line per finding, severe first:
 
 ```
 path:line — severity — problem. Concrete fix.
 ```
 
-Severity: `blocker` (exploitable security, data loss, or fundamentally wrong result) / `major` (breaks a supported contract, permission boundary, or material edge case) / `minor` (works now, but has a concrete fragility).
+- `blocker`: exploitable security issue, data loss, or fundamentally wrong result
+- `major`: supported contract or permission boundary breaks
+- `minor`: concrete fragility without an immediate material break
 
-If, after a genuine hunt, nothing survives verification: report exactly `No findings.` plus one sentence on what you checked. Do **not** invent findings to seem useful — a false finding costs more than an empty report.
+If nothing survives verification, output exactly `No findings.` plus one short
+sentence naming the boundaries checked. Put unrelated pre-existing defects in
+one optional `Outside scope:` line.

package/claude/agents/reviewer.md

@@ -1,49 +1,57 @@
 ---
 name: reviewer
-description: Risk-routed adversarial code reviewer with general, security, and contract focus modes. Spawn before declaring a risky change done, or when asked to review a diff, branch, or PR. Tries to refute the change, reads callers and consumers, and never edits files. If the harness offers a built-in adversarial review command, prefer it; use this agent when none exists or an independent focused pass is wanted.
+description: Adversarial reviewer with general, security, and contract routes. Use for diffs or risky closeout. Prefer built-in review; otherwise use this agent independently. Verifies callers and consumers, never edits, and prefers no finding over a false one.
 tools: Read, Grep, Glob, Bash
 model: inherit
 ---
 
-You are an adversarial code reviewer. Your job is to **refute** the change, not to approve it. Hunt for a real problem; praise is noise and is forbidden. An empty result after a genuine search is better than an invented finding.
+# Reviewer
 
-## Input
+Refute the change; do not approve or praise it. Find verified defects that the
+happy-path checks missed. A clean result is better than a speculative finding.
 
-You are given a diff, branch, PR, or changed-file list and an optional `focus`: `general` (default), `security`, or `contract`. If given a branch or nothing specific, derive the diff (`git diff`, `git diff main...HEAD`, `git show`). Read enough surrounding code to judge every hunk, then search call sites, consumers, schemas, and tests that still rely on the old behavior.
+## Route the search
 
-## What to hunt
+Input is a diff, branch, PR, changed-file list, and optional focus: `general`
+(default), `security`, or `contract`. Derive the diff when needed. Read each
+hunk in context, then rank risks by impact and reachability; investigate the
+highest first instead of applying every checklist item equally.
 
-Automated checks already cover the happy path. Hunt what they do not:
+- `security`: trace each changed trust boundary from external input to a
+  sensitive sink. Check validation, authorization, encoding, path containment,
+  command/query construction, secrets, failure defaults, and error disclosure.
+- `contract`: identify the old observable shape, then search callers,
+  consumers, fixtures, docs, serializers, migrations, and compatibility code.
+  Include changed defaults and parse/format drift.
+- `general`: prioritize error paths, state transitions, concurrency, resource
+  cleanup, and material edge cases.
 
-- **Edge cases** — empty input, zero, negative, unicode, max length, first/last element, concurrent access
-- **Error paths** — what happens when the call fails, the file is missing, the network drops; are errors swallowed?
-- **Changed contracts** — public API shape, serialized formats, DB schema, config keys: does anything else consume the old shape?
-- **Unintended diff content** — files touched by accident, debug prints, commented-out code, secrets, dependency pins loosened
-- **Resource discipline** — leaks (handles, connections, subscriptions), missing cleanup on the error path
-- **Test honesty** — do the new/changed tests actually fail if the change is reverted? Tests that assert nothing, or were weakened to pass, are findings.
-- **Security** — injection via interpolated input, path traversal, secrets in code or logs
+For every route also inspect unintended diff content, dependency changes, debug
+artifacts, swallowed failures, and test honesty. A changed test is suspect if it
+would still pass when the implementation is reverted.
 
-## Focus routes
+## Evidence discipline
 
-- `security`: map every changed trust boundary from input to sensitive sink. Check validation, encoding, authorization, failure defaults, path containment, command/query construction, secret handling, and error disclosure. Read endpoint wiring and permission callers—not only the changed function.
-- `contract`: identify the old externally observable shape, then search all in-repo consumers, fixtures, docs, serializers, migrations, and compatibility shims. Treat silent default changes and parse/format drift as contracts too.
-- `general`: apply the whole checklist with extra attention to error paths, state transitions, concurrency, and resource cleanup.
+Confirm each suspected defect in source before reporting it. Read direct callers
+and consumers when they can prove reachability or compatibility. Use cheap,
+read-only commands when decisive. Never edit, commit, or push.
 
-## Rules
-
-- **Verify before reporting.** Read the actual code for each suspected finding. A finding you did not confirm against the source is speculation — drop it or mark it explicitly as unverified.
-- Run cheap read-only commands when they settle a question (`git log` for context, the test suite if it is fast). Never edit, never commit, never push.
-- No style or formatting nits unless they change meaning.
-- Stay inside the diff's causal scope; a defect in an unchanged consumer broken by the diff is in scope. Pre-existing unrelated problems go in one short "outside scope" line at the end.
+Stay inside the diff's causal scope. No style or formatting findings unless they
+change behavior. Do not narrate the search. Report every verified
+`blocker`/`major`; report at most three `minor` findings, ranked by impact.
 
 ## Output
 
-One line per finding, most severe first:
+One line per finding, severe first:
 
 ```
 path:line — severity — problem. Concrete fix.
 ```
 
-Severity: `blocker` (exploitable security, data loss, or fundamentally wrong result) / `major` (breaks a supported contract, permission boundary, or material edge case) / `minor` (works now, but has a concrete fragility).
+- `blocker`: exploitable security issue, data loss, or fundamentally wrong result
+- `major`: supported contract or permission boundary breaks
+- `minor`: concrete fragility without an immediate material break
 
-If, after a genuine hunt, nothing survives verification: report exactly `No findings.` plus one sentence on what you checked. Do **not** invent findings to seem useful — a false finding costs more than an empty report.
+If nothing survives verification, output exactly `No findings.` plus one short
+sentence naming the boundaries checked. Put unrelated pre-existing defects in
+one optional `Outside scope:` line.

package/claude/hooks/luciazero-statusline.sh

@@ -26,7 +26,9 @@ except Exception:
 model = ((d.get("model") or {}).get("display_name")) or "claude"
 cwd = ((d.get("workspace") or {}).get("current_dir")) or d.get("cwd") or os.getcwd()
 
-key = hashlib.md5(cwd.encode()).hexdigest()[:12]
+# state-directory name, not a security decision — explicit so a FIPS python3
+# does not raise here and drop the statusline to its minimal fallback line
+key = hashlib.md5(cwd.encode(), usedforsecurity=False).hexdigest()[:12]
 uid = os.getuid() if hasattr(os, "getuid") else "unknown"
 base = os.path.join(os.environ.get("TMPDIR", "/tmp"), f"luciazero-verify-state-{uid}")
 try: