[ScriptedAlchemy]

Root cause and scope

Refs #4481 — partial retry-loop repair; not issue acceptance. This PR fixes a distinct LCM convergence redrive path on the session store. It does not assert the original macOS idle CPU target or 24-hour store growth is resolved.

A failed retained-summary invalidation can leave a row in lcm_summary_convergence_dirty_raw (the raw revision remains genuinely dirty). In record_outcome, any such row overrode a retryable, unavailable, or permanent outcome with pending, reset failure_count, and discarded next_attempt_at_ms. The next daemon convergence page therefore immediately retried the same deterministic failure rather than honoring the session's durable retry/parking outcome.

The raw-revision CAS already refuses settlement when a new revision arrives. Keep an old dirty row as pending only for successful/in-progress (current/pending) outcomes. For failed outcomes, preserve the retry deadline or terminal parking until an actual new raw revision increments the generation and rearms the queue through the existing trigger.

Behavioral regression coverage

• Dirty invalidation with a retryable storage failure retains its deadline; a second failure lengthens the deadline rather than hot-looping.
• Reopening the same SQLite store preserves the deadline and failure count.
• A fresh raw revision requeues promptly and rejects settlement from the stale generation.
• unavailable and permanent outcomes park despite outstanding older dirty rows, but a new raw revision unblocks them.
• Successful/in-progress settlement still cannot falsely report complete while dirty invalidation is owed.

Two files only: crates/tracedecay-lcm/src/summary_convergence.rs and its existing crate behavioral tests. No changes to schema, dependencies, policy thresholds, or generated surfaces.

Verification and limits

• Independently reviewed the exact two-file Git object diff before publishing.
• A SQLite predicate reproduction of the old vs fixed settlement shows pending/0/0 before and retryable/1/100000 after for identical dirty input, but this is not a Rust test or macOS daemon benchmark.
• Bazel and macOS daemon are unavailable in the authoring runtime. New Rust tests have not executed here; hosted Bazel CI on the exact PR head must be checked. No cargo command was used.

Separate owners / acceptance

No overlap with #4793 projection batching, #4798 retention GC, or #4799 doctor reporting. No work moved through held draft #4776, Kev/decision-model branches, or release #4763.

#4481 stays open until the full idle-daemon CPU, retry-convergence and 24-hour bounded-growth acceptance is reproduced on master.

[changeset-bot]

⚠️ No Changeset found

Latest commit: ef7d7bf

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