feat: streamline compact reviews with input compatibility proof

[cskwork]

The current compact contract requires multiple independent review contexts, while an assumed input schema can exclude records the original application accepts. Require executable caller/legacy-input evidence before scope freezes and let one independent compact reviewer cover every verification lens and the final adversarial diff on unchanged source. Full-route review remains separate; scenario floors, approval authority, source freshness and retry caps remain required.

The existing Side effects lens now owns domain-specific missing, empty, null, wrong-type and unknown-value probes. Reviewer reports retain actual execution evidence or explicit unknowns, and current proof can be reused only with matching inputs. Windows guidance includes a saved-script preflight executed with spaced paths and process-scoped settings.

The tripwire fix removes confirmed authorization/prohibition false positives while preserving mixed risk and quoted, same-line and multiline removal of credential restrictions. Regression coverage exercises the real approval command; no gate, test or source-integrity check was weakened.

Stacked on open #25.

Validation at candidate f15658f0a7653b14608904f7f36eb8a652bcb1fc:

• Independent review: no source blockers; 14 scanner and 14 real approval cases passed, plus the executed Windows preflight.
• Permanent tripwire regressions: 31 passing observations, including real gate refusal and no approval file for boundary removal.
• Local runner: 7 tests, one existing POSIX-only skip on Windows. Local gate selftest retains the native-symlink permission failure; it is not reported fully green.
• Product lifecycle: final feature 7/7 and bug 8/8, including local review-branch delivery and delivered-tree checks. Hosted Linux/macOS/Windows gate, runner and factory E2E checks all passed in run 38145997479 on this exact candidate.
• Completed matched real-coding comparison: no kit/current kit/revised kit, two previously unseen tasks, three trials, identical requested GPT-5.6-Sol/high/priority settings, 1,200 seconds and three available helpers per run. All 18 outcomes retained, with no scored retries or exclusions. Frozen protocol SHA256: 54f5cb4fb652e46178e0b77ae9a6c2cf92fa12884bd6d7303b151e8761b2e9fb.

Arm	Product checks	Median first observed passing product	Median parent elapsed	Workflow timeouts
No kit	6/6 runs, 48/48 assertions	52.508 s	85.249 s	0
Current kit	6/6 runs, 48/48 assertions	330.000 s	916.375 s	0
Revised kit	6/6 runs, 48/48 assertions	420.000 s	841.563 s	1

First-pass time is the first captured immutable snapshot passing the frozen finite suite, with 15-second capture cadence and last-failed intervals retained; it is not exact completion time or proof of no other bugs. Revised kit's pooled median parent elapsed is 8.2% lower than current kit's, but first pass is 27.3% later. Recovery is descriptively faster; the feature is slower and includes the timeout. Revised kit remains about 9.9× no-kit's median total time, with no measured correctness advantage.

Independent audit: mechanical completion is no-kit 6/6, current-kit 6/6, revised-kit 5/6; useful participant test execution and final test/approval source binding are proved for all 18. Full workflow is six true for no-kit, six unknown for current kit, and five unknown/one false for revised kit. Native helper waits lack child IDs/prompts/responses; written PASS reports cannot authenticate reviewer execution. This interface may conceal real helper activity, so missing receipts are neither zero helpers nor synthetic failures. Four catalog runs replaced a dedicated original export case; replacement coverage may subsume its behavior, so semantic weakening remains unknown and product acceptance is unchanged.

The methods audit preserved all frozen tooling/packages, 90 observations and 1,196 archived file hashes. All runs are serial and cleanup is confirmed. Normal user configuration changed after freeze; actual invocations explicitly ignored user configuration and matched all explicit controls, but complete effective-prompt identity is unproved. Parent-only usage is incomplete for the timeout; helper totals and billing are unavailable. These are descriptive results from three trials per task/arm, not a significance or cost claim.

The user authorized merging validated source fixes to main. These two modest tasks still do not establish that the revised workflow beats baseline; no universal quality, speed, cost or allowance advantage is claimed. The negative comparison and verification limits remain part of this PR.