fix-bot-comment disposition hardening: idempotent retry, fail the run on failed target verification, behavioural test for the node-id guard (#1928 round-3) #1934

Summary

Follow-up hardening for the fix-bot-comment disposition flow added in #1919 (PR #1928). These are round-3 review findings on PR #1928. All three fail safe: the maintainer-comment gate stays blocked, so nothing is wrongly cleared. But each can leave a PR stuck with no clear signal. They were deferred to stop an unbounded review loop on #1928, which had already taken four rounds of fixes.

1. Retry after a partial run posts a second disposition (Codex P1, prompts/dev-lead/fix-bot-comment.md:132)

If a run posts the disposition reply but is cancelled or fails before resolve_dispositioned_comments minimizes the original, the retry finds the original still unresolved and posts a second marker. The resolver requires exactly one authorized disposition and skips otherwise, so the notice stays blocking.

Fix: before posting, look for an existing dev-lead:comment-disposition id=<node_id> reply by the agent and reuse it. Or make the resolver accept multiple identical authorized dispositions for one id.

2. A failed target check doesn't fail the run (Codex P1, prompts/dev-lead/fix-bot-comment.md:127)

When a valid-looking node id cannot be fetched, is the wrong node type, or fails the author check, the prompt says to post nothing and note it in the summary. A failed Bash call inside the agent does not necessarily make the writer engine exit nonzero, and resolve_dispositioned_comments succeeds after resolving zero comments. The run therefore finishes green with the notice still blocking.

Fix: have the harness verify the target itself (before or after the engine), or check afterwards that a disposition for COMMENT_NODE_ID exists and fail the run otherwise.

3. The fail-loud node-id guard is only tested by message text (cubic P3, tests/dev-lead/unit/test_agent_marker_convention.bats:128)

The parse-step guard in .github/workflows/dev-lead-reusable.yml is tested by grepping for its error text. A regression that keeps the message but drops exit 1, or inverts the INTENT_TYPE condition, stays green.

Fix: extract the guard into a small script or function, and add behavioural tests:

• INTENT_TYPE=fix-bot-comment with a missing or malformed id → nonzero exit;
• INTENT_TYPE=fix-bot-comment with a valid id → zero exit, id exported;
• another intent with no id → zero exit.

Rollout

These are dev-lead changes. After merge, cut dev-lead/v139-next and let the canary rings promote.