fix(bin): resolve captain decision holds with multiple dependencies

[jbassiri]

Intent

Fix bin/fm-decision-hold.sh so a captain-decision hold with MULTIPLE dependencies can be resolved. Today such a hold can never close, so it re-surfaces forever and re-pesters the captain with decisions already made. Root cause: tasks-axi quotes a rendered value only when it needs to, so blocked_by renders unquoted for one dependency and quoted ("a,b") for two; show_field kept the surrounding quotes, so the resolve membership test (case ",$blocked,") and the unblock loop tested ,"a,b", and matched no id, failing with 'not durably blocked' against a genuinely present edge. A single-dependency hold passed only by accident, which is why the bug went unnoticed.

Chosen fix (deliberate): the centralized shape - strip one surrounding quote pair once inside show_field itself, which fixes BOTH the resolve path and the unblock loop in one place. This is coupled with a mandatory second change: verify_resolution_identity's resolution_prefix loses its now-stripped leading quote, because the same strip also applies to the escaped body field. Both hunks are required together or the idempotent re-resolve identity path breaks. I deliberately did NOT use the narrower shape that strips at the two command_resolve call sites only, because that leaves the unblock-loop path broken. Scope is strictly the two decision-hold files; salvage commit 011ce5a carried the same fix but also killed turn-end-guard work that was explicitly deliberately excluded.

Test discipline (required by this work): tests/fm-decision-hold-lifecycle.test.sh gains a failing-first regression that was observed RED against the unfixed code (error: 'routed task sample-multi-dep is not durably blocked by ...') before the fix. It builds a single-dep hold (unquoted render), a multi-dep hold (quoted render), and an unrelated blocker; asserts the rendered shapes; resolves the multi-dep routed task; then asserts the hold reaches state done with 'Resolution recorded by fm-decision-hold', the single-dep task is blocked:no, the multi-dep task still blocked:yes carrying ONLY the unrelated edge, and an identical re-resolve is idempotent (exercising the coupled identity path).

What Changed

• show_field in bin/fm-decision-hold.sh now strips one surrounding quote pair from a rendered tasks-axi value. tasks-axi quotes a value only when it needs to, so blocked_by rendered unquoted for one dependency and quoted ("a,b") for two; the resolve membership test (case ",$blocked,") and the unblock loop matched no id against the quoted form and failed with "not durably blocked" against a genuinely present edge, so a multi-dependency hold could never close. Stripping centrally fixes both the resolve path and the unblock loop in one place.
• verify_resolution_identity's resolution_prefix drops its now-stripped leading quote, since the same strip applies to the escaped body field. This hunk is required together with the first one, or the idempotent re-resolve identity path breaks; the escaped newlines stay literal in the rendered value.
• tests/fm-decision-hold-lifecycle.test.sh gains a regression case that builds a single-dep hold (unquoted render), a multi-dep hold (quoted render), and an unrelated blocker; it pins both rendered shapes, resolves the multi-dep routed task, and asserts the hold reaches done with the resolution record, the single-dep task is blocked: no, the multi-dep task still carries only the unrelated edge, and an identical re-resolve is idempotent. docs/decision-hold-lifecycle.md clarifies that resolve clears only each routed task's edge to the hold, leaving other blockers intact.

Risk Assessment

✅ Low: The change is a small, well-bounded two-hunk fix whose safety is provable from tasks-axi's TOON quoting rules (a value containing a quote is always quoted, so the strip cannot over-strip), both coupled hunks are present and consistent, scope matches the stated intent exactly with no unrelated work removed, and it ships a targeted regression test covering the quoted-render resolve, selective unblock, and idempotent re-resolve paths.

Testing

Completed 1 recorded test check.

• Outcome: ⏭️ skipped across 1 run (21m7s)

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

⚠️ **Review** - 1 warning

• ⚠️ bin/fm-decision-hold.sh:121 - show_field strips tasks-axi's surrounding quotes but does not unescape, so command_hold's title equality check at line 268 still mismatches for any title containing a double quote or backslash. tasks-axi quotes AND escapes such values (title: "Choose the &fix(watcher): make check wakes lossless via watcher-side suppression kunchenguid/firstmate#34;north&fix(watcher): make check wakes lossless via watcher-side suppression kunchenguid/firstmate#34; route"), so after the strip the value is Choose the &fix(watcher): make check wakes lossless via watcher-side suppression kunchenguid/firstmate#34;north&fix(watcher): make check wakes lossless via watcher-side suppression kunchenguid/firstmate#34; route and never equals the caller's Choose the "north" route. The idempotent hold retry then fails permanently with 'existing captain hold <id> has a different title' - the same never-closes failure class this change fixes for blocked_by. Not a regression (pre-fix every quoted title mismatched, including ones merely containing a comma or colon; the strip shrinks the failing set), and it is outside the intent's declared resolve/unblock scope, so flagging rather than fixing. Options: unescape only at the title call site, or normalize the comparison. Escapes must stay literal for the body, which verify_resolution_identity parses via literal \n.

⏭️ **Test** - skipped

• 🚨 tests failed with exit code 1
• command -v tmux >/dev/null || { echo "tmux is required for e2e tests" >&2; exit 1; }; tmux -V; rc=0; for t in tests/*.test.sh; do echo "== $t =="; bash "$t" || rc=1; done; exit "$rc"

⚠️ **Document** - 1 info

• ℹ️ docs/decision-hold-lifecycle.md:40 - docs/decision-hold-lifecycle.md's "Verification record" (dated 2026-07-14) hand-copies the per-test ok - transcript of tests/fm-decision-hold-lifecycle.test.sh. It was already incomplete at the base commit (it omits "terminal single-owner stale status decisions do not block empty inventory", which landed in the same commit cd218f2 that wrote the record) and this change adds a ninth case, so it now lags by two lines. I did not edit it: it is presented as the exact output of one dated run, so appending new lines under the old date would falsify it, and re-dating it would require re-running the full 71-script suite plus lint - an out-of-scope rewrite. Follow-up worth considering: either regenerate the block from a fresh run with a new date, or reduce it to the commands plus a pointer to the test files so it stops drifting on every added case.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.