fix(code-index): preserve cancellation epoch after neutral wake

[ScriptedAlchemy]

Problem

At central SHA 65a2610393ac11ca3d2f7dc6452248a49f529109, request_background_reconcile_with_change marks the scheduler dirty for a neutral wake. If the source mutates before the queued reconcile drains, the observed-change wake then sees newly_dirty == false and does not advance the cancellation epoch. The already-running reconcile can consequently publish stale output even though a real source change was observed.

Deterministic reproduction

1. Queue a neutral background reconcile wake.
2. Mutate the source before the queued work drains.
3. Request background reconcile with the observed source change.
4. Assert that the cancellation epoch advances and stale work is cancelled.

Expected

A neutral wake may coalesce scheduling, but it must not consume the transition that records a real observed change. The observed-change request must advance cancellation independently of whether work is already dirty.

Acceptance

• Add a deterministic scheduler test for neutral wake -> source mutation -> observed-change wake before drain.
• Advance cancellation for the observed source change without adding a second authority or changing time/memory budgets.
• Run focused non-vacuous tests and strict relevant clippy.

[ScriptedAlchemy]

Fix is pushed on codex/fix-neutral-wake-cancellation at exact SHA 5e3c79e5b973744697610787eb7f3f8699593ea0 from central floor 65a2610393ac11ca3d2f7dc6452248a49f529109.

Root fix:

• source-neutral overflow and observed-source-change cancellation now use distinct pending markers;
• the observed marker and cancellation epoch advance are serialized under the canonical hint mutex;
• ordinary and retained captures bind one pre-capture control through post-capture validation;
• retained rebuild and unpublished-candidate publication use the same control at the durable-pointer visibility fence;
• unpublished candidates are compared against a fresh authoritative capture, including exact source revision, after current worktree identity is re-resolved and authorized;
• freshness metadata and stat evidence remain bound to the validated capture.

Deterministic RED evidence:

• cc-5155: neutral wake then real edit failed to cancel admitted work.
• cc-5219: edit during capture published the stale pre-edit snapshot.
• cc-5241: retained edit after capture published the stale candidate.
• cc-5308: transient B candidate published after indexed source advanced to C.
• cc-5342: transient B candidate published after a non-indexed C commit moved only source revision.

Exact-tip GREEN evidence:

• cc-5348: real transient publication journey proves B refusal, A preservation, C revision publication, query serving, and zero sealed decode.
• cc-5351, cc-5352, cc-5353, cc-5354, cc-5356, cc-5355: six focused exact tests, each 1 passed and non-vacuous.
• cc-5357: cargo clippy -p tracedecay-code-index-runtime --all-targets -- -D warnings passed.
• cargo fmt check, git diff check, and commitlint passed.

Independent read-only review is CLEAN at exact SHA 5e3c79e5b, tree 4856b24bd35a3dbe10a67c653536f21adbe4c9e1. Local and remote branch SHAs match. Leaving this issue open pending central integration and hosted CI.

[ScriptedAlchemy]

Merged as 95787c3, pushed with 3ac37c2; hauler cc-5990 shows observed_change_after_neutral_wake_supersedes_in_flight_text_work, observed_change_during_capture_retries_before_publication and graph_off_change_after_capture_refuses_stale_publication all PASS in cargo nextest run -p tracedecay-code-index-runtime.