perf(adapter-manifest): validate a manifest once per call, CC-637 batch 1 (guard hook 5.3 s to 3.4 s on Windows)

[screenleon]

Summary

• adapter_manifest_runner_kind, adapter_manifest_effective_route and adapter_manifest_dispatch_path re-ran the full manifest validation (two directory resolutions, two manifest reads) once per nested helper, about five times in dispatch_path. They now validate once and share a new internal _adapter_manifest_runner_kind_of. The checks and their order are unchanged.
• Why it matters beyond tests: guard-executor-write.sh calls these on every codex write; on native Windows one hook call took 5.3 s.
• Records CC-637 (this batch is requirement 1; requirements 2 to 4 are open) and links it to the existing CC-612, which found the same hotspot by profiling (CC-612 is now partial: no cache was added, only the nested repeats were removed).

Evidence (one machine, single measurements, noisy)

• dispatch_path 1757 to 740 ms, effective_route 711 to 399 ms, guard-executor-write.sh call 5.3 to 3.4 s.
• Differential run, old library from main versus new: 4 real adapters and 13 broken manifests (bad schema, name, runner_kind, route, entrypoint, duplicate keys), 4 functions each = 80 outputs, stdout, stderr and exit code byte-identical. A critic subagent ran an independent 97-line version under plain, set -e and set -u: identical.
• test-executor-router.sh: 36 passed, 1 failed; the failure ('missing nonexec directory and symlinks rejected', no symlink or exec-bit support on this host) is identical on main (34 passed, 1 failed there, before the two new cases). test-hook-profile-parity.sh: 5 passed. test-guards.sh --filter exw:: 30 passed.
• Mutation checks: removing the new helper's export -f fails the child-shell case; skipping runner_kind_valid fails the direct invalid-kind case.

Gate and review

• Full pr-gate: NOT run (not feasible on this host); no gate result artifact.
• Five reviewers ran individually as subagents (same model family as the author): critic, security, architecture approve; qa-tester approve with a coverage gap; risk WARN. Risk found a real bug in the first commit (the new helper was missing from the export -f block, so a child bash that inherits the exported public functions failed with command not found); fixed in the second commit with a regression case. QA's surviving mutant (no test for an invalid runner_kind called directly) got a case too.
• Known gaps: the full test-guards.sh was not run (over 10 minutes here, and one background attempt was killed for memory pressure); symlink, exec-bit and boundary checks cannot be exercised on this host, so CI is the evidence for them; test-doctor.sh, test-install.sh and test-pmctl-adapter-generate.sh (partly) not re-run for this PR. adapter_manifest_dispatch_path already fails in a child shell that only inherited the exported functions (its safety patterns do not survive the export round trip); identical on main, not addressed here.
• Permanent test admissions: two cases added to tests/shell/test-executor-router.sh (an existing suite).
  • adapter_manifest: exported functions work in a child shell: stable supported boundary (the library exports its public API for child shells), high-impact regression (a missing export made every public function fail there), implementation-independent observable (exit code and output), low maintenance, direct contract; killed by the missing-export mutant.
  • adapter_manifest_runner_kind: invalid runner_kind is rejected directly: same criteria; the shared validation helper is what the refactor touched and no test reached it directly; killed by the skip-validation mutant.

🤖 Generated with Claude Code

https://claude.ai/code/session_011c6rVDk6zyfgLPEfbVrZi9