refactor: complete explicit session lifetime migration

[thymikee]

Summary

Complete the fixture migration to explicit session publish/update/retire operations and remove SessionStore.set. Retain captured occupant identity and independent successor-fence controls. Depends on #3144. Related to #3116.

Validation

d5472b0523: pnpm check:affected --base 62d6af6340 --run passed. Test Files 329 passed (329); Tests 2177 passed (2177). Selected tooling passed. Rebased onto the corrected optional recorder without expanding this layer. Fresh GitHub CI remains pending; ancestor native evidence and approval are separate requirements. User handles merges.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.98 MB	4.98 MB	-137 B
Package (unpacked)	4.98 MB	4.98 MB	-137 B
Package (download)	1.49 MB	1.49 MB	-25 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.6 ms	26.5 ms	-0.1 ms
CLI --help	81.2 ms	79.6 ms	-1.6 ms

[cubic-dev-ai]

No issues found across 100 files

_{Re-trigger cubic}

[thymikee]

This PR is ready. I reviewed commit 26f32c6 and found no problems in the test and fixture changes. The diff touches no production files, and all 13 checks are green. There are no conflicts. It can merge once the base PR #3151 lands.

Not blocking: the comment above the "moved to another device" test in daemon-session-idle-expiry-scheduling.test.ts (https://github.com/callstack/agent-device/blob/26f32c6/src/daemon/server/daemon-session-idle-expiry-scheduling.test.ts#L309) still says a command can replace the record with one bound to a different device, but the fixture now patches the same lifetime with sessionStore.update, so rewording it would help. You can take or leave this.

I did not run the test suites. I took the pass counts in the PR body and the green checks as reported. I did not read all 100 files line by line. I read the timer-based retire tests, the update-based rebuilds, the fixture rewrites and the removed re-sets, and I checked the mechanical set-to-publish swaps by count only. Whether each publish swap is safe on an occupied address rests on the passing suites, since publish throws session_address_occupied there.

[thymikee]

The three new commits since a8430eb look good at 1bec0d3. They only touch test fixtures and finish the removal of SessionStore.set and delete, so I found no code problems in the delta.

CI is still pending. All 9 non-passing checks are queued, with no failures shown yet. Lint & Format, Coverage, Integration and Smoke would exercise this code, so please wait for them to finish green. Typecheck is the final proof that no SessionStore.set or delete callers remain.

I did not run typecheck or tests locally. I also did not verify your note that the earlier iOS smoke failure was a daemon-startup registration-busy exit. If Smoke fails again, please share the run so we can confirm it is unrelated.

The head moved to 3766ea1 after this review. The new commit comes from the base branch (it joins daemon dispatches before shutdown retires session admission), so this verdict still describes 1bec0d3. I will check the new head separately.

[thymikee]

This PR is ready at a94493c. The earlier review at a8430eb (#3152 (comment)) was clean, and the changes since then are test and fixture only, with no production code touched.

Not blocking: the R7 gate header in scripts/layering/session-state.ts (lines 2 and 12) and the header in scripts/layering/session-resource-ownership.ts (line 4) still give SessionStore.get()/set() as the reason modules can mutate store-owned state, but set() is gone and writes now go through update(ref, patch), publish and retire, so you could reword those three comments to say get() returns the live record and writes use that API, or leave them.

Smoke Tests and Bundle Size show as cancelled from a superseded run, not failed, and this delta does not touch anything either job exercises. A fresh run on a94493c should settle both, and Typecheck must also finish green there; typecheck is what proves no SessionStore.set or delete caller remains, and I did not run it or the tests, only grepped the production code at this head. I did not re-check the earlier iOS smoke failure that was tied to a daemon-startup registration race, since it sits outside this delta. Cubic and Copilot have not reviewed this head, and there are no open review threads. There are no conflicts, so the only thing left before merge is those green runs on a94493c.

[thymikee]

The PR is ready. The earlier clean review at a94493c still holds, and at d5472b0 the rebased migration onto the new base adds no new problems.

Not blocking: the comment at session-replay-repair-loop.test.ts line 283 still says the earlier waits are "on the discarded session", but no session is discarded now, because the openRebuildsActions option clears the action list on the same record, so you could reword it to say open rebuilt the action list (and note the scenario is synthetic, since production open keeps actions), or leave it as is.

CI is still pending: Smoke Tests is running with no failure so far, and the other checks are green. I did not run tests locally, so the claim that the rebased migration compiles and passes against the new base rests on CI. I did not re-review the full diff outside this update, which the earlier clean reviews covered. The only route-relevant production change here is the alert and settings recording through recordSessionAction, and it comes from base PR #3144, not from this PR. No conflicts. Before merge, Smoke Tests must finish green and #3144 must merge first.