S07 part 1: plan grounding check (observe-only, flag off)

[im-tyler]

Changes

• src/plan-grounding.ts: pure extractor and judge plus a git-backed adapter. It checks whether the files, symbols, package.json scripts and make targets a plan names exist in the committed tree at a stated revision (git ls-tree, git show, git grep against the commit, so the working tree is never consulted).
• Each reference is grounded, ungrounded, unchecked (could not look) or proposed (the plan says it creates it). unchecked is never counted as a pass.
• A mismatch between the expected revision and the tree asserts nothing: every reference is unchecked.
• Flag: SHIP_PLAN_GROUNDING=on|1|true (default off). Nothing is wired into any run, route or durable step, and storage is unchanged. No model is called.
• src/plan-grounding.test.ts: 15 tests over real temp git repos read through a real LocalExecutor.

Checks

• pnpm run lint is clean. Root suite is 1998 pass, 0 fail (1983 before plus 15 new), run via node --test --experimental-test-isolation=none "dist/**/*.test.js". After merging main I re-ran lint, build and the new test file only.
• The grader-sensitivity script was skipped (fixed port 8901 clashes between parallel agents). CI runs it.
• Negative controls, each mutated, shown failing, then restored:
  • search the working tree instead of the revision: 2 tests fail
  • report a truncated listing as ungrounded: 1 fails
  • ignore a revision mismatch: 1 fails
  • ignore the "creates" marker: 1 fails
  • list files from the working tree instead of the commit: 2 fail
  • report an unverified command as grounded: 1 fails
  • remove the symbol-grammar guard in the adapter: 0 failed at first, because the extractor grammar already blocks it. I added a direct adapter test, and it then fails (1).

Not verified / still open

• "Grounded" means the name exists, not that the plan is correct. Symbols are matched by identifier text, not by semantics.
• Commands other than root package scripts and root make targets are unchecked (nothing is run).
• The advisory is not shown on the plan review view yet (see the S07 parts that remain, below).
• The rest of S07 is not done:
  • the web advisory display
  • revision-provenance citations
  • inference-versus-reproduction labels
  • plan coverage of acceptance criteria
  • the coding-loop diagnostics
  • budgeted alternative candidates
  • answer-correctness and attribution grading

🤖 Generated with Claude Code

https://claude.ai/code/session_01VqsBNqvaWezf1DwQrAnVgX

---

Generated by Claude Code