test(llm): repair confidence hierarchy lookup

TODO.md

@@ -2,6 +2,11 @@
 
 ## Active tickets
 
+- [ ] [`ticket-030`](project/ticket-030/README.md) — repair the stale
+  confidence-hierarchy source lookup after the NL and Markdown clamp logic
+  moved into focused helper modules. Current state: `IN_PROGRESS / VALIDATION`;
+  the production ceilings remain unchanged and the implementation is test-only.
+
 - [ ] [`ticket-019`](project/ticket-019/README.md) — publish the dependency-free
   Python SDK as the root PyPI distribution `todo2code` through `goal -a`, with
   one root `pyproject.toml` and SDK-only artifacts. Current state:

project/TICKETS.md

@@ -27,4 +27,5 @@ of `project/README.md`, which remains a generated technical-analysis artifact.
 | **ticket-019** | [`README.md`](./ticket-019/README.md) | [`preprompt.md`](./ticket-019/preprompt.md) | - |  [`ai-codex.md`](./ticket-019/ai-codex.md) |  [`ai-codex-logs.txt`](./ticket-019/ai-codex-logs.txt) | [`changelog.md`](./ticket-019/changelog.md) |
 | **ticket-020** | [`README.md`](./ticket-020/README.md) | [`preprompt.md`](./ticket-020/preprompt.md) | - |  [`ai-codex.md`](./ticket-020/ai-codex.md) |  [`ai-codex-logs.txt`](./ticket-020/ai-codex-logs.txt) | [`changelog.md`](./ticket-020/changelog.md) |
 | **ticket-022** | [`README.md`](./ticket-022/README.md) | [`preprompt.md`](./ticket-022/preprompt.md) | - |  [`ai-codex.md`](./ticket-022/ai-codex.md) |  [`ai-codex-logs.txt`](./ticket-022/ai-codex-logs.txt) | [`changelog.md`](./ticket-022/changelog.md) |
+| **ticket-030** | [`README.md`](./ticket-030/README.md) | [`preprompt.md`](./ticket-030/preprompt.md) | - |  [`ai-codex.md`](./ticket-030/ai-codex.md) |  [`ai-codex-logs.txt`](./ticket-030/ai-codex-logs.txt) | [`changelog.md`](./ticket-030/changelog.md) |
 <!-- AUTO:TICKET_INDEX:END -->

project/ticket-030/README.md

@@ -0,0 +1,71 @@
+# Ticket 030: Repair confidence hierarchy test after extractor refactor
+
+- **ID**: ticket-030
+- **Owner**: unresolved:human
+- **Status**: IN_PROGRESS
+- **Workflow state**: VALIDATION
+- **Created**: 2026-08-04
+
+## Goal and scope
+
+Repair the confidence-hierarchy regression test that failed after the Markdown
+and NL confidence clamping logic moved from the orchestration modules into
+focused helper modules. The production ceilings remain correct (`0.94`,
+`0.90`, `0.85`); the failure is a test-discovery defect, not evidence that an
+Intent record contains `confidence: null`.
+
+The implementation is deliberately test-only. It will update the hierarchy
+test so it verifies the canonical current locations without changing any
+extractor behavior, confidence value, documentation contract or LLM boundary.
+No unrelated refactor or generated artifact belongs to this ticket.
+
+## Planned changed paths
+
+- `test/nl-llm.test.ts`: replace the stale source-file assumptions with a
+  deterministic check of the current confidence-clamping implementations.
+- `project/ticket-030/**`, `TODO.md`, `project/TICKETS.md`: ticket intent and
+  verification evidence only.
+
+## Acceptance criteria
+
+- [x] AC-01: Scope is approved by a human owner.
+- [x] AC-02: The test distinguishes a missing source-code match from a runtime
+      confidence value and no longer reports the former as extractor failure.
+- [x] AC-03: The hierarchy remains Markdown `0.94` > NL `0.90` > documentation
+      `0.85`, with all three ceilings below deterministic observation levels.
+- [x] AC-04: Focused NL/Markdown/document tests pass without a live provider,
+      changing production code or accepting a looser confidence ceiling.
+- [x] AC-05: The ticket diff contains only its allowed test and governance
+      paths; unrelated dirty/generated files remain untouched.
+
+## Participants
+
+- Human participant: unresolved; no user-* file was created by this script.
+- Agent participant: [ai-codex.md](ai-codex.md)
+
+## Approval boundary
+
+- Current state: `IN_PROGRESS / VALIDATION`.
+- Required response from: `unresolved:human`.
+- Chat approval authorizes implementation in this interactive session but is
+  not trusted merge evidence.
+- The ticket was renumbered from the locally drafted `ticket-023` to
+  `ticket-030` before publication because remote PR #3 already owns 023. The
+  approved implementation scope and code delta did not change.
+
+## Validation evidence
+
+- The freshly emitted `dist/test/nl-llm.test.js` passes `11/11`; the repaired
+  hierarchy assertion passes and no live provider is called.
+- The reported 338-test run used an older `dist` tree: it correctly exposed the
+  stale hierarchy lookup, but by itself was not proof that current TypeScript
+  sources build cleanly. After rebasing onto base `caf6551`, a clean
+  `npm run build` passes.
+- The full freshly built suite reports 338 tests: 327 passed, 10 failed and one
+  skipped. The repaired confidence-hierarchy test passes. The ten inherited
+  failures concern AST/linker relations, communication identity/enrichment and
+  gold-dataset relations; none is in ticket-030 scope or caused by its two-line
+  source-location update.
+- Scoped `git diff --check` passes. Whole-repository governance continues to
+  report only the pre-existing ticket-018/ticket-019 conflict, dependency,
+  workstream ownership and shared `Makefile` overlap.

project/ticket-030/ai-codex-logs.txt

@@ -0,0 +1,8 @@
+2026-08-04 ticket-030 validation
+- `node --test dist/test/nl-llm.test.js`: PASS (11/11).
+- `npm run build`: FAIL (17 inherited TypeScript diagnostics outside allowed implementation path).
+- `npm test` after the failed fresh emit: FAIL (292 total, 253 pass, 38 fail, 1 skip); base import/runtime regressions exposed.
+- The user-reported 338-test run was based on stale `dist`; useful for reproducing the hierarchy assertion only, not for clean-source verification.
+- `make governance`: FAIL with the four inherited ticket-018/ticket-019 diagnostics (`GOV-CONFLICT-001`, `GOV-DEPENDENCY-002`, `GOV-WORKSTREAM-003`, `GOV-WORKSTREAM-004`); no ticket-030 diagnostic was reported.
+- Rebased onto updated PR-5 base `caf6551`: `npm run build` PASS.
+- Fresh `npm test`: FAIL (338 total, 327 pass, 10 fail, 1 skip). Failures are AST/linker relation (2 direct plus gold cohorts), communication identity/enrichment (3), and configuration/document relation coverage; the ticket-030 test passes.

project/ticket-030/ai-codex.md

@@ -0,0 +1,56 @@
+---
+participant-id: agent:codex
+participant: codex
+role: agent
+ticket: ticket-030
+---
+# Participant: codex (AI agent)
+
+## Understanding
+
+The failed assertion's `actual: null` is the result of
+`source.match(...)`, not the confidence stored in an Intent record. The test
+still reads `markdown-llm.ts` and `nl-llm.ts`, while their clamping expressions
+now live in `markdown-llm-helpers.ts` and `nl-llm-helpers.ts`. The documented
+and runtime ceilings remain `0.94`, `0.90` and `0.85`.
+
+The narrow repair belongs to the `llm` workstream and does not overlap the
+active governance or SDK implementation paths. It must not absorb the user's
+concurrent synthesis refactor or generated `project/` artifacts.
+
+## Execution plan
+
+1. Stop at `WAIT_FOR_APPROVAL` before editing the test.
+2. Update only the confidence-hierarchy test to inspect the canonical current
+   implementations, retaining strict numeric assertions and ordering.
+3. Run the focused compiled/TypeScript tests without a live OpenRouter call.
+4. Review the exact diff and run scoped governance; record any inherited
+   repository failures separately from this test-only change.
+
+## Actual changes
+
+- The user explicitly approved the bounded ticket-030 plan in chat. The
+  implementation session is authorized; merge approval remains external.
+- Updated only `test/nl-llm.test.ts`: the hierarchy test now reads the Markdown
+  and NL helper modules that own their confidence clamps after refactoring.
+- Ran the freshly emitted compiled test file: 11 passed, 0 failed, with the
+  hierarchy assertion passing. No provider request was made.
+- Rechecked the updated base from a clean dependency installation. Base
+  `caf6551` builds successfully, and the freshly built complete suite reports
+  327 passed, 10 failed and one skipped out of 338. The confidence-hierarchy
+  test passes; the remaining failures are inherited relation and communication
+  regressions outside this ticket.
+- Renumbered the unpublished local ticket from 023 to 030 after discovering
+  that GitHub PR #3 already owns ticket-023; no implementation path or
+  acceptance criterion changed.
+
+## Blockers
+
+- Whole-repository governance already fails because ticket-018 and ticket-019
+  declare a conflict and unfinished dependency. This ticket does not rewrite
+  either intent or claim that unrelated blocker.
+- The main worktree contains user-owned generated and synthesis changes. Work
+  remains isolated in a separate branch/worktree.
+- Full verification is blocked by ten source-behaviour test failures outside
+  the approved ticket paths. The base branch needs separately scoped repairs
+  before it can be considered merge-ready.

project/ticket-030/changelog.md

@@ -0,0 +1,16 @@
+# Ticket Changelog (ticket-030)
+
+## [0.1.0] - 2026-08-04
+
+- Initial governance scaffold created.
+- No human participant identity or content was generated.
+
+## [0.1.1] - 2026-08-04
+
+- Repaired the stale Markdown/NL confidence-clamp source locations in the
+  hierarchy test without changing production confidence behavior.
+- Focused compiled NL LLM test passes 11/11. After rebasing onto `caf6551`, a
+  clean build passes; the complete fresh suite exposes ten inherited relation
+  and communication failures while the repaired hierarchy assertion stays
+  green.
+- Renumbered the unpublished ticket to 030 to avoid colliding with active PR #3.

project/ticket-030/intent.json

@@ -0,0 +1,17 @@
+{
+  "schema": "new-project.intent/v2",
+  "ticket": "ticket-030",
+  "summary": "Repair confidence hierarchy test after extractor refactor",
+  "workstream": "llm",
+  "allowedPaths": [
+    "test/nl-llm.test.ts",
+    "project/ticket-030/**",
+    "TODO.md",
+    "project/TICKETS.md"
+  ],
+  "forbiddenPaths": ["project/ticket-*/user-*.md"],
+  "stacks": ["node"],
+  "dependsOn": [],
+  "conflictsWith": [],
+  "integrationTicket": null
+}

project/ticket-030/preprompt.md

@@ -0,0 +1,8 @@
+# Ticket preprompt
+
+- **Task ID**: ticket-030
+- **Task title**: Repair confidence hierarchy test after extractor refactor
+- **Created**: 2026-08-04T09:30:20Z
+
+Keep executable implementation outside this governance/evidence directory.
+Read a human-owned user-*.md file only when one exists.

test/nl-llm.test.ts

@@ -192,8 +192,8 @@ test('The documented confidence hierarchy holds across LLM extractors', async ()
   // Read the TypeScript sources from the repository root: the compiled test
   // runs out of dist/, where the .ts files do not exist.
   const sources = await Promise.all([
-    fs.readFile(path.resolve('src/extractors/markdown-llm.ts'), 'utf8'),
-    fs.readFile(path.resolve('src/extractors/nl-llm.ts'), 'utf8'),
+    fs.readFile(path.resolve('src/extractors/markdown-llm-helpers.ts'), 'utf8'),
+    fs.readFile(path.resolve('src/extractors/nl-llm-helpers.ts'), 'utf8'),
     fs.readFile(path.resolve('src/extractors/docs-record.ts'), 'utf8'),
   ]);
   const ceilings = sources.map((source) => {