Extract StepTraceSession and port the step-cache tests against it

[kblok]

The step cache logic ran inline in the act, so the tests could only reach it through a whole session and could not check what upstream checks: the host reads in order, the staged recording, and the exact miss reason. It now lives in StepTraceSession (src/E2E/Agent/StepCache.cs) with StepOutcome and FailedStepOutcome, and the act calls it as upstream's act does. The step-cache.test.ts tests are ported against that seam with a MakeHost helper; the through-the-pipeline versions stay under _through_the_pipeline names. Behavior does not change, except that CANCELLED now keeps the cache like a model outage, as upstream's failedStepOutcome does. The built-in agent's start capture now falls back to a second observe when the session's probe of the start screen fails, so a flaky first capture still lets the step run but records nothing.

.NET-only symbols: IStepCacheHost.FirstActedOn (the port reads the delta from the screen the first action was resolved against), StepCacheOptions with the cache context folded in, ClaimedKey, the private WatchedHost, StepTraceSession.Pending/Start, the REPLAY_STALE case in FailedStepOutcome, and the test types StepCacheHarness, StepHost, CacheContext, StubCache. ReplayOptions.LooksBeforeFree is upstream's looksBeforeFree and replaces the host's OnEndRoute. The port keeps its own gap/strict rules and settles evictions in Flush, not in conclude.

Partial tests (renamed): "keys on the params with each unique() value as a placeholder..." (the key claim and slotting are the attempt's and dispatcher's; kept as Records_nothing_and_says_so_when_a_unique_value_is_spelled_by_another_param for the collision half), "looks at the page a navigate opened before free actions" (..._before_a_free_press...: no untargeted typing), "stages the entry it replayed whole to keep" (no recorded verdict), "self-finalizes when the recorded destination arrives after the replayed action" (summary wording), "still runs live a step with no recording and a recording too long to replay" (no truncated half), "leaves an entry the same flow re-recorded untouched" (no summary or end wait to vary; kept as the pipeline test). Not ported, feature missing: recordedVerdictOf (an entry records no verdict), the three pixel-evidence tests, truncated entry, no start anchor, the two end-wait tests, live progress replaying, "leaves an entry untouched when only the rule differs", "says a confirmed recording ... was not written", and "compares the step as the entry stores it" (the act masks the step before the session).

Closes #382

🤖 Generated with Claude Code

https://claude.ai/code/session_01KWUdrjfarTEM4dmBbK6GaV