[Agent tooling] worktree_guard follow-ups from the PR #1851 review: submodule fixture, path flavour, strict expectations, ADR note

[Chris0Jeky]

Residuals from the PR #1851 review (worktree_guard substance validation, #1833):

1. Substance check 1 (linked-worktree gitdir containment) has no failing test. Deleting the whole case "$_wt_gitdir" in "$_wt_worktrees_dir"/* block leaves the suite green (35/35) — every rejected fixture is caught by check 2 alone. The isolating case exists: a git submodule work tree has a .git pointer file resolving exactly to its own git dir, so only check 1 rejects it. Add a submodule fixture so the check has a test that can fail.
2. WT_GIT_DIR is exported in MSYS flavour (/c/...) while WT_REPO_ROOT is Windows flavour (C:/...) in the sh guard under Git for Windows (also visible in the FATAL diagnostic). Nothing consumes WT_GIT_DIR today; it is a trap for the first consumer that hands it to a native Windows tool. Normalize or document.
3. Contradictory HEAD expectations resolve silently. WT_EXPECT_HEAD=detached + WT_EXPECT_BRANCH=foo (and the ps1 equivalents) ignores the branch instead of erroring. A setup-error exit (2) would be more fail-closed.
4. expect_sh_output asserts only the NOTE substring, never the exit code — a guard that printed the NOTE then failed would still pass that case. Cheap to tighten.
5. Test-harness robustness: if both mktemp -d forms fail, FIXTURE_ROOT is empty and fixtures land at /primary, /bare.git, ... Guard the mktemp result.
6. docs/WORKTREE_AGENT_PROTOCOL.md:14 is stale — "Other roots are allowed only when the coordinator explicitly names them and updates guard configuration" no longer matches the substance-validated guard. (Coordinator docs sweep may fix this line first; keep the box here until it lands.)
7. Consider a short ADR recording the boundary change (path-shape markers demoted to advisory; substance validation now the gate). Repo CLAUDE.md names automation safety boundaries as ADR triggers; the net effect is strictly stronger, which is why PR fix(worktree): validate worktree guards by substance, not path shape #1851 shipped without one.

Source: PR #1851 fresh-context review (non-blocking items 1–7). None merge-relevant.

[Chris0Jeky]

Claiming bounded lane alpha-product-trust / TD-1855-exit-status from base 9cf28bbc60925344c77a8cdaed9ad8cc37064634. Scope: tighten expect_sh_output so the advisory NOTE case also requires the shell guard to exit 0, add a synthetic NOTE-then-failure mutation control, and align the two stale common-directory expectations with the guard's documented fatal exit 1 contract. Owned path: scripts/worktree_guard.tests.sh only. Submodule containment, fixture-root failure, path normalization, contradictory expectations, protocol docs, and ADR decisions remain out of scope.

[Chris0Jeky]

Released bounded slice TD-1855-exit-status in PR #2642: #2642

• commit a47b0bcb3671af46e0ff7695df2be907403fe288
• expect_sh_output now requires exit 0 as well as the advisory NOTE
• synthetic NOTE-then-failure canary added
• stale common-directory pointer expectations corrected to the documented fatal exit 1 contract
• final self-tests: 36 passed, 0 failed, 0 skipped; bash -n and git diff --check passed
• two coordinator review passes found no blocking findings

Automatic Codex review is unavailable because the account review quota is exhausted. No merge performed.

[Chris0Jeky]

Evidence-only receipt: TD-1855-protocol-note

Verified on origin/main at exact head 9eb6f404513df55828bc46c5f03c4537fd59e111: docs/WORKTREE_AGENT_PROTOCOL.md already describes substance-based linked-worktree admission, accepts unconventional roots with an advisory NOTE, and states that path shape is not the safety property. The stale sentence identified by this row is already corrected; no docs PR is needed.

The remaining #1855 rows are separate test/code/decision lanes, including the open PR #2642.

Refs #1855

[Chris0Jeky]

Coordinator note (2026-09-05, D-12 sweep): the Codex session's PR on this issue was reviewed by one fresh-context Opus 5 reviewer (record on the PR) and merged with a merge commit. The PR body used Refs, so this issue stays open; what merged and what remains, from the review:

• PR #2642 merged as 67ab09a70: expect_sh_output requires exit 0 as well as the NOTE, a NOTE-then-failure canary proves output alone cannot pass, and the two stale common-directory pointer expectations now match the guard's documented exit 1. LOWs tracked here: the suite runs in no workflow (only the three PowerShell harnesses run in reusable-docs-governance.yml), which is why the two expectations sat red unnoticed; the rewritten comment attributes the malformed-pointer exit 2 to the guard's layout check when it comes from git failing repository discovery; the canary swallows subshell diagnostics on failure. PowerShell implementation, submodule and path-normalization residuals remain.