fix(cache): delete both breakpoint spellings from the split's volatile half

[OsherElhadad]

Closes nothing on its own — this is a blocker found by integration review of the five cache/filter/observe merges (#33, #36, #40, #42, #43). Neither per-PR review could see it.

The defect

splitVolatileTail copies the system block it splits, then deletes cache_control from the churning half so the breakpoint covers only the stable text (apply/prefixsplit.go:118).

It did not delete cachePoint — the Bedrock/Vertex spelling of the same thing, which wireBreakpoints counts alongside cache_control (apply/metawrite.go:117).

So on those providers, a splittable system block carrying an inline cachePoint gets its breakpoint duplicated — one on each half, 1 → 2. That silently spends a slot from the provider's cap of four, and from a request already at the cap it puts 5 on the wire → 400.

Why both reviews missed it

The two halves landed in the same PR (#36) but in different files: the split predates it, the cachePoint counting was added by it. And the Anthropic path — the only one the benchmarks exercise — has no cachePoint to duplicate, so no run could have surfaced it.

This is the same shape as the bug #36 was written to fix (46 breakpoints applied, 0 forwarded): a provider-specific wire detail invisible to the code path under test.

Fix

One line, plus a comment explaining why both spellings must go.

delete(volatile, "cache_control")
delete(volatile, "cachePoint")

Test

TestSplitDoesNotDuplicateBedrockCachePoint asserts the wire breakpoint count is unchanged across the split, and that the volatile half keeps no breakpoint in either spelling.

Verified non-vacuous — with the fix reverted it fails with the exact reported symptom:

--- FAIL: TestSplitDoesNotDuplicateBedrockCachePoint
    prefixsplit_test.go:353: split changed the wire breakpoint count 1 -> 2;
    a copied cachePoint burns a slot from the provider's cap of four

Gates

go build -tags cg_skeleton ./... · go test -tags cg_skeleton ./... · go test -race ./apply/... · gofmt -l · go vet — all clean.

Related findings from the same review, NOT fixed here

Tracked separately so this stays a one-line blocker fix:

1. The volatile-tail split is a silent no-op on Bedrock Converse. explicitBreakpointProvider admits Bedrock/Vertex, but on real Converse shape cachePoint is its own array entry after the block it terminates — so the split inserts the volatile half before it and the breakpoint still covers the churn. The split reports Changed: true and cachesplit looks active in /stats, yet the −34.1% win it is credited with cannot materialize there. Same class of defect, one provider over.
2. Observe mode leaks hypotheticals into unlabelled /stats fields. cmdfilter's FilterStats sink has no Mode check, so an observe-only run reports real cmdfilter_families/cmdfilter_filters/cmdfilter_selector_misses with no mode label and no potential_* counterpart — a consumer cannot tell enforced from hypothetical. New leakage from feat(cmdfilter): port rtk's filter coverage (3 -> 24) and fix three DSL loss-typing bugs #42 × feat(proxy): add observe mode — measure what compaction would save, without enforcing it #43 composing.
3. Pin-budget exhaustion can reopen the TailOnly fail-open on the /compact route only (the chat path uses modes.Tracker and never consults cg:len:). Measurement-accuracy risk for offline eval, not live traffic.
4. The selector-miss ledger bounds count, not key size — replaying multimodal traffic filled the top slots with base64 PNG blobs, making the "which filter to write next" ledger unusable on that traffic.