fix(gate): replay a script gate's evals, and name its script by file not address

[jothimani-rajendran]

What

Two things a policy shipping a kind: script gate (0.9.3) hits the moment it is adopted from a catalog. Found by moving java-security in open-coder-ai/chock-catalog onto the script gate against the v0.9.3 pin.

1. A script gate's evals could not pass. chock check --only evals replays a staged-files case against the compiled gate in a throwaway repository holding only the case's own files. That is whole for a declarative gate, whose JSON is the entire check, and empty-handed for a script gate, whose program lives under the policy's implementations/ and is resolved from the repository root: every case observed script gate: '...' is not installed -- refusing and blocked, so a suite could only pass by expecting block (25 cases: 13 false passes, 12 fails). execute.py now copies the policy's implementations/ to where the compiled gate names the script, as chock sync would have, before running the case. Copied, never staged: the material a case puts before the gate stays the files it lists, and a test pins that a script which would refuse its own text does not refuse every case.
2. A script gate's packaged SKILL.md differed by address. The ambient line rendered the compiled script param, the file's path from the repository root, which is base/<id>/implementations/x.py in a catalog tree and .agents/policies/<id>/implementations/x.py once adopted. chock plugin build --check therefore reported plugin_drift in the adopter for a package that was clean in the catalog. advisory.py renders the bare file name, which is what the manifest declared, so one packaged skill is right in both places.

Docs: spec/gate-dsl.md (both behaviours), CHANGELOG.md under Unreleased.

Definition of done

• chock check → 0 errors, 0 warnings (the 5 infos are pre-existing on main)
• chock check --only matrix passes (32 invariants); no behavior change to the matrix
• chock sync --repo . --check clean
• chock check --only verify clean
• Registry rescanned; no stale entries
• pytest -q green: 1428 passed, 2 skipped (26 script-kind tests, 6 new: replay allow/block through run_case, the script is never in the writes, and the ambient line is identical for the policy at .agents/policies/ and under a catalog tree)
• pytest acceptance/ -c acceptance/pytest.ini --rootdir=acceptance green: 21 passed
• Existing artifacts: no policy in this repo uses script; nothing to migrate
• Touched manifests: none. The stability-script golden's ambient-rule/ambient.md is regenerated with CHOCK_REGEN_GOLDENS=1 (one line: the script param); no other golden changed
• ruff check . and ruff format --check . clean

Emitter output changes for the script kind's ambient line, so under CONTRIBUTING's stability rule the release carrying this is a minor one, even though the kind shipped yesterday and nothing published uses it yet. A maintainer may reasonably call it a patch on that ground; the golden regeneration is in this PR either way.

Claims

• No surface is described as enforcing more than it installs. INSTALLED_SURFACES, the coverage table and docs/enforcement-surfaces.md are untouched: neither change alters what is emitted or installed, only what an eval replays against and what one rendered line says.

🤖 Generated with Claude Code

---

Generated by Claude Code