fix(core): restore sessions and propagate errors on session/within/retryTo error paths

[DavertMik]

What

Implements the actual fixes for the latent promise-core divergences that #5622 characterized. Each fix makes an error path symmetric with its already-working success path, and the corresponding characterization assertion is updated to the corrected behavior in the same commit.

⚠️ Stacked on #5622 (base = advisor/001-promise-core-characterization); review only the fix(core): … commit. Merge #5622 first.

Fixes (4 of the 7 findings)

#	Finding	Fix
5	within() async error skipped _withinEnd and detached the error	catch now runs finishHelpers() (so _withinEnd runs on error) and return recorder.promise() (so the error propagates through within())
3	session() async error leaked the session id	catch now schedules recorder.session.restore via recorder.add so the recorder session is restored on error
4	session() sync queued-throw leaked the session id	the finally recorder.catch now calls recorder.session.restore before re-throwing (the only place that runs on a rejected chain)
6	retryTo() rejected prematurely / tries === 2 on first	a thrown callback routes through recorder.throw instead of an immediate reject(err), so the retry logic owns the outcome — throw-then-succeed now resolves; reject only after maxTries. tries starts at 1; retry count preserved (tries < maxTries)

Important: the session fix is deliberately surgical

The session error paths are entangled with the Playwright browser lifecycle. An earlier version of this PR routed the async-error cleanup through finalize() — which also swaps the original (no-op) restoreVars(sessionName) call for the real restoreVars(). That closed the browser context under BROWSER_RESTART=session and failed the Playwright CI job (4 Cannot create page: browser context has been closed failures in session_test.js).

The fix is now minimal: only add the missing recorder.session.restore (the session-id leak the finding is about); the existing restoreVars/listener cleanup is left exactly as-is. Net lib/session.js change is two added lines.

Verification

• npm run test:unit → 748 passed, 0 failed, 11 skipped
• npm run test:runner → 273 passed, 0 failed — incl. should rerun flaky tests, should rerun retried steps, should use retryFailedStep plugin for failed steps, etc.
• Acceptance (real Chromium), both restart modes — BROWSER_RESTART=browser and BROWSER_RESTART=session — within/session/els green, 0 browser context has been closed failures. (The one remaining local acceptance failure is config_test.js's API call to the local mock at :3001/api/users, which 404s only in this sandbox and passes in CI — unrelated to these changes.)

Deliberately not changed here

1. recorder.retries never cleared by reset() — lib/plugin/retryFailedStep.js (on by default) reads/mutates recorder.retries; the in-code comment warns clearing it breaks the plugin. Needs per-test scoping coordinated with that plugin.
2. Stopped recorder → silent hang — add()'s no-op-when-stopped is a contract the should not add steps when stopped test depends on; these error-path fixes reduce the wedging, but a blanket change is unsafe.
3. Nested cross-level restore order — low severity; real within/session keep start+restore at one level. The characterization test pins the correct synchronous id semantics.

Happy to tackle any of these three as a focused follow-up.

🤖 Generated with Claude Code