[agent] Filed by the scheduled architecture audit (audit-cli-tests). Register: register comment.

Kind: refactor (test honesty). Register ID: T62.

Problem

The docker agent suites (docker_e2e_<eco>.rs) are the only end-to-end proof that scan → scan --sync → apply works against a real installed package. Eight of the nine serve a placeholder "beforeHash": "000…0" from their wiremock fixture and then run apply --force. For example, golang has the placeholder at L98 and the force apply at L194.

The same pattern, a 64-zero beforeHash plus --force, appears in cargo, composer, deno, golang, maven, npm, nuget and pypi (grep-counted on cbbbdb9).

gem is the exception, and its own doc explains why the pattern is a problem (docker_e2e_gem.rs L88-L95):

Serving this as the fixture's beforeHash — instead of an all-zeros placeholder — is what lets the gated apply paths run: with the placeholder, scan --sync's nested apply hit the variant gate and failed invisibly, so the chain only ever proved the --force path.

As a result, in the other eight suites the nested apply that scan --sync performs is expected to fail. The suites log its exit code instead of gating on it (golang L180: "Exit code is logged for…"). The gated, non-force apply path (beforeHash match → write → afterHash verify) has no container-level proof for those ecosystems.

Proposed change

1. Generalise gem's upstream_before_hash probe, a once-per-process probe container that computes the real git-blob sha256 of the installed file, into the shared docker harness (see Share one docker test harness: cov_docker_args ×10, skip_if_no_image ×11 and assert_vex_agent_attested ×8 are copied, and 8 docker suites never honour SOCKET_PATCH_DOCKER_E2E_REQUIRED #1968).
2. Serve the real hash in each suite's mock.
3. Gate on the nested scan --sync apply's exit code and summary.applied.
4. Keep at most one --force leg per suite, and name it as a force test.

Do one ecosystem per PR, or all eight at once if CI time allows.

Size

8 test files. About +15 / −10 lines each, plus about 40 shared lines.

Acceptance criteria

• In each suite, the non-force chain passes in coverage-docker with scan --sync reporting applied: 1.
• Setting the served beforeHash back to zeros makes the non-force leg fail.
• Any remaining force legs are named *_force_*.

Related: #1968 (T60).