[Haoming9527]

Refs #872. Stacked on #945 (work/872-scan-metadata).

What was wrong

Every admitted PTC loader callback assembled the entire public prompt again to obtain a presentation boolean. With ten callbacks in one step, the real host fixture performed 12 assemblies (11 probes). Native/discovery callbacks and result admission also repeated this inspection.

The change

Cache the final public probe verdict by prompt service and exact agent identity. Owned public prompt/tool change listeners invalidate all verdicts; disposal removes reuse, unobserved deployments fall back to a fresh probe, and invalidated asynchronous probes cannot authorize execution or overwrite a newer verdict. Callbacks retain live scope/definition/grant checks and compare the revision across asynchronous admission. No private executor API, dependency, schema, output limit or permission contract changes.

Includes bilingual documentation, changelog, three cache regressions, a real-host ten-callback regression, reproducible benchmark and every warmup/measured sample with source hashes.

Validation

• npm run build: passed.
• Presentation/precheck/PTC/discovery/projection/file-publication filters: 17 suites passed, 0 failed, 0 skipped.
• Publication/authority/persistent admission/mode/shared-source/inheritance/result-storage filters: 14 suites passed, 0 failed, 0 skipped, with the real PR perf(search): reuse directory-relative metadata snapshots #945 worker.
• Behavioral red/green: the new precheck regression on the before checkout fails with 11 probes versus expected 1 (2 other checks pass); candidate passes all 3 checks, including rejection after a same-step executor change.
• git diff --check and source-size gate: passed.
• 300 measured steps after 30 retained warmups per arm; ten SDK loader callbacks per step. Identical results across 3,000 calls in each arm; retained callbacks refused after executor switch in both arms.

Measurement	Before p50 / p95	After p50 / p95
Loader callback	13.39 / 18.12 ms	0.42 / 0.80 ms
Step preparation	15.49 / 22.28 ms	15.61 / 20.24 ms
Preparation + ten callbacks	153.17 / 181.22 ms	20.31 / 25.79 ms
Total assemblies per step	12	2

This is a publication benchmark with real host services and a resident fixture that rejects execution. It excludes worker work, run_code compilation, shell and model requests. The full 106-operation scenario, full orchestration SLO and shell performance acceptance remain unmeasured or unmet. This PR does not close #872 or authorize a runtime release.

Pull-request-triggered CI: https://github.com/T-Auto/dsh-ops/actions/runs/43613610235 — all 10 jobs passed, including CI required and the Windows/Linux/macOS full release gates. The first run exposed a stale test assumption: ptc.test waited for an assembly despite a valid cache. The race fixture now explicitly invalidates the cache and fails if execution bypasses its barrier. The full PTC suite passes locally with normal staged-worker resolution; the assertion that an executor change prevents worker dispatch remains intact.

Environment

Windows x64 10.0.26300, Ryzen 7 6800HS, Node 26.12.1, locked dependencies. Base #945 at 9a626c16fe64e092bf25b8269073d15248e1628c; before host source is #944 at 24ac360551c1336453b78268090574f97a1de50a (unchanged by #945). No merge, tag or release.