chore: Snowflake Cortex caching follow-ups (walk-back continuation, image-block probe, cross-process cooldown)

[anandgupta42]

Consolidated follow-ups from the #1020 consensus code review and the v0.9.2 release review. Deferred because: each needs design or live verification beyond a quick fix.

1. attachMarker walk-back: when the target block already has a marker, continue scanning earlier blocks/messages for an unmarked spot to preserve a second trailing breakpoint (MiMo/GLM-5.1 finding; rare in the shipped pipeline).
2. Live-probe whether Cortex honors cache_control on image blocks; add a unit test for a newly-relocated marker landing on a non-text last block.
3. Cache-disable cooldown (cacheDisabledUntil) is per-process; concurrent processes each pay the probe-retry tax independently on marker-rejecting accounts (Chaos Gremlin P1). Consider shared state only if such accounts are ever observed.
4. Discussion: should a (non-SyntaxError) transform bug be able to fail a chat turn, or degrade to pass-through with a loud log (Tech Lead P2)?
5. Dormant paths (src/plugin/snowflake-cortex.ts, packages/core cortexFetch PluginV2 port) lack the caching fix — port it if either is revived.