fix(code-index): isolate workers and keep the reader charge

[ScriptedAlchemy]

Cause

Strict readiness on PR dogfood (tip 3f71bccb2dba, run 35341667443) sat in bulk_commit with graph=pending until the 600s deadline. The first pass was also unscheduled for 226s on the daemon runtime (queue_delay_micros=226558274, wake_to_gates_held_micros=15).

A follow-up that started graph activation while that text projection was still in flight inverted the parked-owner ordering. Those convergence tests fail when an in-flight projection is treated as graph admission. That overlap is not in this branch.

CI run 35442762975 then failed on the rebased head: workspace clippy (u64::try_from of Metadata::len in the sessions JSONL test), and search_never_awaits_an_in_flight_decode_while_a_generation_is_servable timing out at 360s then passing in 0.3s in a fresh process. Test Linux only echoed that partition failure.

Fix

• The daemon polls code-index workers on their own runtime, so project-open work cannot hold the first pass. Tests spawn on the test runtime instead. A process-wide runtime of a few threads is the only poller for every parallel mount, and a waiter sits behind the others until the per-test deadline.
• Both publication tails hand the charge they already hold to code-text-artifact-reader and shrink it to the reader budget. They do not drop it and reserve again. The clone-successor charge is at least the reader budget.
• Graph seating still waits for a ready text owner.
• Metadata::len is used directly. It is already u64.

Rebased onto master 993f5e3c9bce4cc2a03dc43ac0621f1f6ab85819. ci.yml is unchanged.

Proof

code_index_scheduler::tests::serving::search_never_awaits_an_in_flight_decode_while_a_generation_is_servable — 1 passed in 0.56s.

cargo clippy -p tracedecay-sessions -p tracedecay-code-index-runtime --all-targets --locked -- -D warnings — passed.

SHA: fac0a0a475405b78a0691673db4c3d3085a1de59

  

[changeset-bot]

⚠️ No Changeset found

Latest commit: fac0a0a

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-19T07:04:57.183683Z	d323531	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d323531935

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ScriptedAlchemy]

Review lane verdict: REJECT as-is. The graph/text overlap change (mount.rs: text_projection_in_flight treated as 'text owner admitted') inverts the convergence-park invariant two existing tests assert: convergence_park_tests::fresh_graph_activation_waits_while_the_published_text_owner_is_parked and ::fresh_graph_activation_starts_while_the_clone_successor_is_pending fail deterministically (5/5) with this PR's mount.rs/serving.rs and pass with master's. The PR's test plan never ran the crate's --lib suite. Also: clippy -D warnings fails on the added test_gates.rs (question_mark, redundant_closure_for_method_calls); the reservation shrink is applied to only one of the two publication tails (advance_clone_successor still drops and re-reserves); the retained reservation keeps the build component key. The dedicated worker runtime and the reservation shrink (applied to both tails, keeping the reader key) are separable and worth landing as their own PRs; keep the overlap out until the #1103 parked-owner invariant is re-argued with those two tests updated deliberately. Note #1836 stamps a continuation from published_text_projection_outcome before drop(reconcile_pass); under this PR that is None at that point.

[ScriptedAlchemy]

Second review against the head fac0a0a: the overlap change is already reverted here (946ad3b), the shrink is on both tails and keeps the reader key, and the clippy items are fixed. The reservation shrink is now ported onto master as its own PR with a ledger-level test that discriminates transfer from fresh admission. The dedicated worker runtime stays out: it compiles under cfg(not(test)) so nothing can fail without it, its own commit notes it hit the 360 s deadline when many registries shared few threads, and the worker already offloads heavy work through spawn_blocking, so a starved worker points at whatever blocks the shared runtime rather than at the worker. This PR closes once the port merges.

[cursor]

Closing as superseded by #1873.

#1873 is still open but already owns the reservation-shrink port (both publication tails + transfer_component + ledger test). The only unique leftover on this branch is the dedicated worker runtime under cfg(not(test)), which Zack explicitly kept out of the merge path (360s starvation notes; spawn_blocking already offloads). The sessions clippy fix already landed on master via cb730a980c.

Prefer close-as-superseded over a zombie PR that only carries rejected worker-runtime delta. Once #1873 merges, nothing from #1840 remains to land.