review: tighten comments, fix stale claims-journal docs

[CMGS]

Whole-repo /code /code-rs /code-py comment-budget pass plus a docs-staleness check against the recently merged #12–#16. Comment-only + docs; no behavior change. A 5-agent scan flagged ~50 candidates; most were already-tight, load-bearing WHY comments (#7 tightened them) and were deliberately left alone — this PR applies only genuine reductions and two real staleness fixes.

Staleness fixes (the substantive part)

• docs/performance.md — "Claims-journal write path": the section stated "keep the simple in-mutex write; do not move to a store-owned sequential writer." F1 (perf(pool): move the claims-journal write off the manager mutex (F1) #6) and F1-part-2 (perf(pool): marshal the claims journal off the manager mutex (F1 part 2) #14) did exactly that. Rewritten for the snapshot() (field-copy under the manager mutex) / commit() (marshal + write off it) split, with the stale BenchmarkStoreSaveScaling table replaced by a BenchmarkStorePersistContention description (kept to the robust ~6× ratio; absolute ns not re-cited without a fresh hardware run).
• sandboxd/pool/claims.go — claimStore doc: still said "the manager mutex covers only the marshal". perf(pool): marshal the claims journal off the manager mutex (F1 part 2) #14 moved the marshal off the mutex (snapshot now takes a cheap field-copy). Corrected + tightened.

Comment reductions (restate / narrative only)

• sandboxd/server/server.go redirectClaim: dropped scenario-painting, kept the redirect targets + no_redirect ping-pong WHY.
• sandboxd/engine/portconn.go guestPortConn: dropped the {op:data}/{type:data} restate-of-code, kept the wire format + purpose.
• boot/init/src/cfg.rs: removed three test sub-case labels the asserts already state.
• sdk/openai/.../adapter.py: removed one restate comment (kept the load-bearing # 404-safe in the SDK).

Deliberately NOT changed

• Dense multi-line WHY in refill.go, claim.go, pool.go, hibernate.go, mesh.go, relay.go, silkd exec.rs/session.rs/lsp.rs — accurate and load-bearing; re-compressing review: tighten comments + unsafe SAFETY note #7-approved prose would be churn.
• docs/benchmarks.md fs_pull row is stale post-perf(sdk): fast-path bulk frame decode (2.15x exec throughput) #13 (~2× faster) but the log is append-only; a fresh row needs a bare-metal re-bench (.79 currently off), so no fabricated number was added.

Gates

• Go (sandboxd): go build ✓, go vet darwin+linux ✓, golangci-lint run darwin+linux → 0 issues, golangci-lint fmt --diff clean, go test -race pool/engine/server → ok.
• Rust (boot/init): cargo fmt --check ✓, cargo clippy --all-targets -D warnings ✓, cargo test parse_ip_param → ok.
• Python (sdk/openai): ruff check → all passed, py_compile ✓; pytest blocked by the external agents (openai-agents) SDK absent in this env (comment-only change).