L01-02: define complete Core inspection receipt

[mberrys]

Closes #16

Change

• Add a typed Core inspection receipt bound to input digest, revision, effective profile, check coverage, evidence refs, fidelity, limitations, and the canonical four-state verdict.
• Derive a stable identity from input and effective policy; fail closed on missing or unsupported coverage and mismatched provenance.
• Keep the preflight report and evidence-bundle schemas unchanged. Fix the mapped CTest command for multi-config Release builds.

Proof

• Exact source and remote-head SHA: 874c953bd45f989c7881ad08b5480be9cf9c1dfc; base origin/dev: 7ba2361a3cb3489f93a82a45ca96f2755121a543.
• python scripts/agent/check-change.py --base origin/dev --build-dir build — PASS, 72/72 checks. Its mapped CTest run passed 54/54 tests on Windows Release.
• python -m unittest scripts.agent.test_check_change -q — 21 passed. python -m unittest scripts.agent.test_architecture_contracts -q — 8 passed, 1 skipped.
• Golden vectors in UnitTests/tst_preflightverdicttest.cpp: synthetic input digest a × 64 and profile digest b × 64 yield receipt ID 8e94a7fefac162eafa4a20516692fb281c28978788b9d91d4df3397061933a8b; profile digest c × 64 yields 931952b4c72f56e67fa371c94eb01bec4383cab251286a3c675c7d8dcada11f8. The same fixture covers PASS, FAIL, INCOMPLETE, ERROR, missing coverage, unsupported and budget-limited status, cancellation, parser error, and mismatched provenance.
• Hosted checks already green at this SHA include policy, architecture, source integrity, package scripts, CodeQL, C++/Python/Actions analysis, and Semgrep. agent-fast / build and fuzz are still pending; the workflow's linux and windows jobs were skipped. This PR does not claim the L01 parent release-admission gate.

Handoff

Changed: Core verdict header/implementation, verdict unit test, verdict docs, one changelog fragment and evidence manifest, and the mapped gate script. The observed result is a stable typed receipt and fail-closed terminal outcomes with the exact-SHA local proof above. Remaining risk: callers must adopt the in-memory receipt API; scheduled-result fencing and cross-platform release admission remain in L01-03 and L01-07.

Anti-slop review: Reused the existing Core reducer and evidence types, and kept the change in their owning module. Retained only provenance and completion guards needed for a trustworthy receipt; no generated wrappers or broad exception handling were added.