fix(llm): stop rolling cache breakpoints from forfeiting marker slots

[devlikebear]

Summary

Review follow-up to #933 (#921). Cleanup and one real logic gap; placement rules are unchanged.

• Marker slots were silently forfeited. The budget was truncated to the newest two completed turns before markability was known, so a newest turn that cannot carry a marker (empty assistant content, or a block array ending in tool_use) cost a slot outright and the request shipped with one breakpoint or none. Slots are now filled newest-first and consumed only when a marker actually lands, so an unmarkable turn falls through to an older completed turn.
• The budget arithmetic was unreachable. hasSystemBlocks/hasTools could only ever yield 2 — all four combinations collapse through the cap — and the budget <= 0 guard was dead. Replaced by a single reservedBreakpoints count plus anthropicMessageCacheBudget, whose branches are now reachable and pinned per reservation level.
• Two identical branches in anthropicCompletedTurnEndIndexes, plus the prevGroupStart variable that existed only to choose between them.
• Four copies of the cache_control literal, now behind anthropicEphemeralCacheControl() (returns a fresh map per call — callers store it into blocks they do not own).
• Two wrong doc comments. The fallback marker sits one completed turn back, not two. And the in-flight exclusion was justified by "would write a fresh entry per loop iteration without ever being read", which is not true here: agent.Loop appends each tool exchange to the same slice and re-sends it, so a marker on a completed in-flight tool_result would be read by the next iteration. The comment now states this is a deliberately conservative choice and names the in-flight marker as the next step.
• CHANGELOG entries feat(llm): add rolling cache breakpoints to anthropic message history #933 never wrote.

Deliberately out of scope

Actually placing a marker inside the in-flight group, and the Tools: nil final call at internal/agent/loop.go:431 (tools precede messages in the cached prefix, so that call sits in its own cache lineage and pays a 1.25x write). Both change token spend rather than clean up code, and #933's test plan left the cache_read_input_tokens measurement unchecked — they belong in a PR that measures first.

Test plan

• 6 new tests: rolling window advancing across turns, fallback to an older turn when the newest is unmarkable, per-reservation budget table, length-mismatch guard
• Fallback test verified to fail against the old logic ([3] vs [1,3])
• make test — full suite green
• make vet, gofmt clean
• make lint-diff — 0 issues
• make test-cover-diff — 95.7% (45/47 changed lines), threshold 80%

🤖 Generated with Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
94.6% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud