AUDIT-codex-2026-08-11.md gains a closure table: every finding mapped to
fixed / rejected / accepted-as-risk with the commit that did it, plus the two
findings discovered while writing the regression tests and a note on the two
test attempts that mutation checking proved were vacuous.

README and config.example.yaml document the new settings (server.read_timeout,
cache.max_bytes, queue_size: 0 meaning no waiting room) and the new metric
(mcp_gateway_cache_bytes) and label folding. The design-decisions section now
explains three-valued breaker outcomes and why the breaker is acquired after
admission rather than before, and what 'degraded' means now that a skipped
backend counts.

Five items are recorded in Limitations rather than fixed, each with its
reason: calls carrying _meta are never cached; plaintext tokens remain in the
retained config; timeouts apply on the Run path only, not to an embedder
serving Handler() from its own http.Server; and a wedged stdio write is
bounded for callers but its writer goroutine can still park where the platform
gives pipes no write deadline.

Benchmarks re-run after all fixes; changes are within noise and the numbers
in the README are updated to the fresh run.