[LE0-Lin]

Summary

• Add a runnable offline design lab explaining why score-delta-only checks differ from baseline-governed regression gating. Five explicitly synthetic development scenarios exercise the actual module CLI under three defined gate settings (15 subprocess comparisons).
• Include two passing controls, 100-to-100 removal/reversal examples, and a deliberately weakened caller passing candidate thresholds as explicit overrides. Findings remain visible in all modes; this is not a default-tool vulnerability, historical-version experiment or an AI comparison.
• Retain actual observed exits/scores/error IDs, portable commands, runtime source identity and runner/scenario hashes. Do not claim real-world precision/recall, independent ground truth, user adoption, timing savings or agent execution enforcement.
• Add 12 executable contracts and an English lab guide with source/evidence/coverage limits. Link it from the docs/examples index and synchronize the optional Simplified/Traditional user-guide explanations.
• Runtime implementation, scoring, package version, published artifacts, Action refs and frozen evaluation records are unchanged. No external maintainer contact or application-author biography is added to the project.

Verification

• Full Windows suite: 544 tests, 532 passed and 12 platform skips.
• Linux lab suite: all 12 passed, replaying the same observations.
• All 16 existing installed-wheel acceptance journeys still passed; the installed runtime source identity is unchanged (bca1d4976d359c2f6e37a7068c077691461d8fc117d3df6a6ab3fdb5a10f82ef). CI rebuilds/tests the distribution as usual.
• The retained five-scenario table came from an actual local run, not invented measurements. Evidence tier remains synthetic_development_demo; independent accuracy is null, AI comparison false, adoption established false.
• Exact-head PR CI must pass before merge. No new runtime rule or accuracy result is introduced.