fix(grug): editing a PR body no longer throws away a finished Elder review

[quadseven]

Why

Editing a PR body stalls Elder on the same commit.

Elder's pre-publish freshness guard compared only the freshness id. That id also hashes title and body, so a body edit moved it while the code under review was byte-identical, and the guard discarded the finished review as stale_snapshot.

That discard is not one wasted pass. It re-enqueues against the SAME sha, elder_in_progress_check_skipped: already_in_progress then declines to post a check-run because one is still open, and the check sits pending.

Hit live on quadseven/infra#2133 today, and the irony is the point:

03:00:27  code_reviewer_dispatched   head=a5098494  result=pass
03:08     author edits the body to satisfy Grug - Chief's DoR gate
03:08:11  elder_review_enqueued      head=a5098494  snapshot v1:fc9b1775 -> v1:5ff36c54
03:12:31  code_reviewer_dispatched   head=a5098494  result=publish_failed
03:25:23  code_reviewer_dispatched   head=a5098494  result=pass

Same head_sha throughout. A docs-only PR sat pending for 25 minutes because satisfying one grug check invalidated another.

Summary

• Head sha decides staleness. When the reviewed sha is still current, the review publishes - the findings anchor to identical lines and only the intent blurb the model also saw has changed. Logged as code_review_intent_drift_publishing_anyway.
• A genuine push (sha actually moved) still discards as before.
• The cancel path no longer says "superseded by a newer commit" when the watcher fired on a title/body edit and no commit landed.

The trade-off is explicit and not new: review_freshness_id already documents choosing bounded staleness over never publishing, for base_sha. This applies the same reasoning to the other two fields it hashes. A rewritten intent can legitimately change findings, so the re-enqueued pass still runs - it just no longer destroys a good verdict first.

Acceptance criteria

• A freshness-id change with an unchanged head_sha no longer discards the review ([grug] Elder deadlocks: same SHA + changed snapshot_id read as "superseded by a newer commit", cancels then re-enqueues forever #773 criterion 1) - test_intent_only_edit_does_not_discard_a_review_of_the_same_commit
• Repro fixture: in-flight review plus intent change, no new commit ([grug] Elder deadlocks: same SHA + changed snapshot_id read as "superseded by a newer commit", cancels then re-enqueues forever #773 criterion 4) - test_durable_dispatch_publishes_when_only_intent_moved_during_inference drives the full dispatch path
• A real head change still discards - test_durable_dispatch_rejects_stale_input_before_inference, fixture corrected (it computed a changed id but returned the OLD sha as current, so it asserted the bug)
• Cancel message names the real cause
• Full webhook suite green: 1868 passed, 2 skipped, 0 failed

Size

Size: S

Out of scope

• [grug] Elder deadlocks: same SHA + changed snapshot_id read as "superseded by a newer commit", cancels then re-enqueues forever #773 criterion 3, bounding repeated re-enqueues on an unchanged sha so they terminate with a real conclusion instead of pending. This PR removes the main trigger but does not add the backstop.
• [grug] Elder deadlocks: same SHA + changed snapshot_id read as "superseded by a newer commit", cancels then re-enqueues forever #773 item 4, whether llm_backend_transport_failed backend=cave contention is causing the mid-flight cancels or merely co-occurring.
• The ConnectError storm that made the 03:12 attempt publish_failed on infra#2133. That is [grug] Durable retry for deep-review publication failures (post-model transport errors) #738, durable retry for publication failures.

Refs #773