fix(cache): clamp old snowflakes + poller channel backoff

[bakeb7j0]

Summary

Fixes a bug where getMessages returned [] when after was older than the cache window (e.g., after=0), breaking all clients using the standard "give me everything" cursor pattern. Also adds exponential backoff to the poller so channels that consistently fail stop wasting rate-limit budget.

Changes

• cache.ts — Clamp after to window start instead of returning empty; uses - 1n to avoid off-by-one at the boundary
• poller.ts — Add ChannelHealth tracker with exponential backoff (pollInterval × 2^(failures-1), cap 16×). Self-healing: success resets backoff. recordSuccess only reports recovery when channel was in active backoff (not expired).
• index.ts — Wire shared health tracker into initialPoll and startPollingLoop
• tests/ — 9 new tests (cache clamp, after=0, channel health unit + integration), 2 existing tests updated for corrected behavior

Linked Issues

Closes #12

Test Plan

• ./scripts/ci/validate.sh — lint (tsc + shellcheck) + 63 tests passing
• Cache clamp verified: old snowflakes and after=0 return all cached messages
• Backoff verified: failed channels skipped on subsequent polls, recovery clears state
• Code review: off-by-one fixed, false "recovered" log fixed, cap test strengthened