fix(observability): first cache write after below-threshold calls is first_write, not miss_avoidable

[philmerrell]

Problem

The prompt-cache classifier (PR #697) produces a false positive: in a session where every prior call was uncached (prompt below the model's minimum cacheable prefix — e.g. ~4096 tokens for Claude Haiku 4.5), the first call that crosses the threshold does cacheWrite>0 / cacheRead=0 within the TTL window and gets classified miss_avoidable. That inflates the AvoidableMiss and WastedUsd EMF metrics and the admin Session Cost Anatomy page.

Verified live in dev-ai session 9a1f25b2-8ea4-487f-b0e7-80a60deff945: calls 1–5 uncached at 3.5–4k tokens, call 6 falsely flagged miss_avoidable with write=4122/read=0, calls 7–8 clean hits.

Fix

classify_cache_status gains an optional previous_cached_prefix_tokens parameter (the previous call's cacheRead + cacheWrite total, which _derive_cache_observability already computed for waste pricing). When the immediately preceding call had zero cache activity, there was no cache entry to read from — so the threshold-crossing write is classified first_write (the expected initial population) and excluded from waste pricing.

Reusing first_write (rather than adding a new status) keeps the EMF schema, dashboard, and admin anatomy page enums unchanged. None (unknown split) preserves the previous behavior.

Tests

• Classifier: below-threshold-then-crossing → first_write; unknown previous prefix stays miss_avoidable.
• Derivation: previous row with zero cache activity → first_write, wastedUsd == 0.0.
• Full backend suite: 4771 passed, 3 skipped.

🤖 Generated with Claude Code