fix(provider): keep the Anthropic cache breakpoint off ephemeral turns

[TheGreatAxios]

Summary

• The Anthropic Messages cache breakpoint lands on the last persisted non-assistant turn instead of the ephemeral tail, so director nudges no longer bust the cached prefix.
• Covers the stock anthropic provider and both first-party Messages adapters (zen-messages, opencode-go-messages); system and tools breakpoints stay untouched.

Verification

• bun run check passes: lint, typecheck, build, and guard with 7309 tests and 0 failures.
• New src/provider/anthropic-cache-breakpoint.test.ts pins the breakpoint placement for all three providers, system/tools preservation, byte-identical passthrough without ephemeral turns, and non-Anthropic passthrough.

Fixes CL-6913

[linear-code]

CL-6913

[TheGreatAxios]

Critic review · no blocking findings

Moves the Anthropic messages cache breakpoint off the ephemeral nudge tail onto the last persisted non-assistant message, via a registry decorator wrapping the three Anthropic-messages providers.

Verified with evidence

• src/provider/anthropic-cache-breakpoint.ts:23-27 — suffix counts non-system ephemeral turns only, matching the vendored adapter (vendor/intx-inference/src/providers/anthropic.ts:69-70), which extracts system turns out of messages into body.system. System-role ephemeral can never skew the split.
• src/provider/anthropic-cache-breakpoint.ts:42-55 — backward scan skips assistant, mirroring the vendor's isLastUser rule (anthropic.ts:90-93). Confirmed parity by test: persisted [user, assistant] + nudge puts the breakpoint on index 0, identical to the vendor building persisted-alone.
• src/provider/anthropic-cache-breakpoint.ts:33-40,71-74 — strip covers the full suffix, placement targets pre-split only; system/tools wire keys are never touched (only body.messages is parsed/mutated). PR test pins both breakpoints surviving.
• src/provider/anthropic-cache-breakpoint.ts:62,72,74 — fail-open is safe and leak-free: suffix === 0 returns built untouched (byte-identical passthrough, pinned by test); split <= 0 and placement-failure return the original built, discarding the parsed mutation — no half-stripped body ever ships.
• src/provider/inference-dependencies.ts:99 — decorator applies outside withReplaySanitizer, so the split is computed from post-sanitizer wire length against the ephemeral count. Persisted-side drops by the sanitizer cannot desync it; only the ephemeral tail must survive 1:1, and it does (producers emit user turns only; sanitizer drops assistant turns only).
• src/provider/anthropic-cache-breakpoint.ts:89-98 — contract is sync AdapterRegistry -> AdapterRegistry, non-Anthropic providers return the inner adapter by reference. No async drift, no signature drift.
• Position-vs-marks: the decorator trusts count-from-tail rather than verifying turn identity. Sound here — vendor/intx-inference/src/reactor.ts:730-733 always appends ephemeralTurns at the prompt tail, and every producer (src/agent/director.ts, src/subagent/nudge-director.ts) emits user-role nudges. No caller passes ephemeral mid-list.
• Temp hypothesis tests (since removed): two stacked nudges, assistant-ended persisted tail, all-ephemeral prompt, system-role ephemeral — all 4 pass; PR's own 6 tests pass.

Notes

• Keeper tests worth adding (routed, not implemented here): stacked multi-nudge breakpoint placement, and the all-ephemeral fail-open byte-identity case. The fail-open branches (split <= 0, unparseable body, no eligible persisted message) are currently unpinned.
• File-for-later, not this diff: a direct buildRequest caller passing ephemeral turns anywhere but the tail would mis-split. No such caller exists today; if one ever does, the decorator should verify tail identity instead of counting.

No blocking or should-fix findings. Hygiene is clean: two focused files, no dead code, no untouched-file churn, red-then-green commit shape intact.