[AndresDiagoM]

⚠️ base-red inherited: #18791

Summary

This is option 2 from #18692, following the split suggested there. #18819 added the guard, and this PR adds the light survey on top of it. Everything is under tests/, in one new file: tests/unit/executor-frozen-body.test.ts (547 lines). It contains one test, every executor's first model per format runs on a deep-frozen body inside the egress guard.

• It enumerates executors as the harness from question: an "executor contract" test that execute() never writes its input (would you want it?) #18692 did: registry aliases, PROVIDERS, the dashboard catalog and three dynamic *-compatible-* nodes, deduplicated. Ids that refuse chat are skipped and counted.
• For each executor it takes only the first model per request format (openai, openai-responses, claude, gemini). It uses one generic credential profile and one body per format.
• Each case runs once on a deep-frozen body, with globalThis.fetch stubbed. The run happens inside the tests/_setup/egressGuard.ts guard from test(setup): opt-in egress guard for in-process tests, plus the #16631 Claude OAuth frozen-body case #18819, and executors are also constructed inside it.
• It has no pins: no unreachable list, no partial-reach list and no table of known writers.

What fails and what only reports

The test fails only when the test itself has stopped working:

• The guard is still installed after the run.
• A corpus body was not actually frozen.
• For one of the four formats, no executor reached the stubbed fetch. That means the enumeration or the stub broke, and without this check the test would pass without checking anything.

The rest goes to t.diagnostic and never fails the test:

• Writers: the executor, the format, the error and the first repo stack frame.
• Unreached cases, grouped by reason: HTTP status of the local error response, which transport the guard blocked, or the thrown message.

When a case runs past 5 s it is aborted. A timed-out run that ignores the abort could otherwise reach the network after the survey, once the guard and the real fetch are back. So the survey waits up to 10 s for such runs to settle. While any is still running, the guard stays installed and fetch stays stubbed for the rest of the test process, and the test reports it as a diagnostic.

What it reports today on release/v3.8.53

379 executor/format cases (+20 ids that refuse chat), 334 reached the stubbed fetch: openai 305, openai-responses 10, claude 16, gemini 3
WRITES: pollinations/openai          'model'               open-sse/executors/pollinations.ts:60
WRITES: opencode-zen/openai          tool 'name'           open-sse/translator/helpers/toolCallHelper.ts:78
WRITES: opencode-zen/openai-responses adds 'summary'       open-sse/executors/opencode.ts:218
WRITES: opencode-go/openai           tool 'name'           open-sse/translator/helpers/toolCallHelper.ts:78
WRITES: opencode-go/openai-responses adds 'summary'        open-sse/executors/opencode.ts:218
WRITES: vertex/claude                adds 'anthropic_version' open-sse/executors/vertex.ts:477
WRITES: nvidia/openai                tool_calls[].'id'     open-sse/executors/default.ts:154

These are the in-place writers already listed in #18692. As you asked there, I'm not filing them as bugs; the test only lists them. Of the 45 cases that don't reach the stub, 7 are those writers, whose frozen write throws first. The other 38 are web and session executors that stop with a local HTTP error, either because the generic stub can't answer their session bootstrap or because they reject the body (for example a forced tool_choice), CLI and http2/wreq-js transports the guard blocks, and one credential that needs a userId. The output lists them by reason.

Verification

• Five runs (three before the linger change, two after) gave identical diagnostics. The test takes about 4 to 6 s, and about 10 to 16 s wall time including node start.
• Run together with egress-guard.test.ts and claude-oauth-frozen-body-egress-guard-16631.test.ts: 14/14 pass, with DATA_DIR isolated.
• I made three temporary changes to check the test isn't empty. Each was reverted, with git diff --quiet clean afterwards:
  • An in-place write added to MoonshotExecutor.transformRequest: the test still passes, and the diagnostic now names moonshot/openai and kimi/openai at moonshot.ts:145.
  • The stub's reach detection broken on purpose: every format reaches 0, and the test fails with no executor sent a openai request to the stubbed fetch.
  • moonshot made to hang and ignore the abort: the test passes and reports still running 10000ms after their abort, egress left blocked: moonshot.
• ESLint with complexity 15 and sonarjs/cognitive-complexity 15 is clean, and so are prettier and the repo's ESLint run. The file is under the 1440-line cap.
• I did not run the full npm run test:unit locally: on this Windows machine it takes about 3 hours. CI is the full run.

Left out of the full harness on purpose

• The other two credential profiles. The Claude OAuth block stays covered by the two *16631* tests.
• Six of the seven body shapes, the Vertex transport variants and the opencode/muse-spark special ids.
• The second pass on a plain copy, diffed against a snapshot.
• Every pin table.

Refs #18692.