test: drive 402 through real escalation consumers

[thegeorgepu]

What this changes

Replaces the hand-copied 402 control-flow assertion with a deterministic integration test that drives real status: 402 errors through advisor, council, scope, and loopbreak. Each case asserts a terminal throw, exactly one stronger-class attempt, an untouched same-class retry and cheaper fallback, and no unavailable-class or continuation warning. The test is registered in the aggregate Vinci harness.

The test also accepts VINCI_TEST_402_CONSUMER=advisor|council|scope|loopbreak so mutation testing can prove every real consumer fails independently when account errors are misclassified.

Why

Fixes #40. The classifier was pinned, but none of its escalation consumers were; a consumer could silently reintroduce downgrade-on-billing-error while the existing copied test stayed green.

Layer

• vinci/ - the Vinci layer
• packages/**/vinci-*.ts - a Vinci change inside an upstream package
• Something else in packages/ - this probably belongs upstream, see CONTRIBUTING.md
• Docs, CI, or tooling only

Exact successor

• Base main: 6e19f1418718e9e33fdfa2d0d6c03054c5768470
• Head: 240ddebe1948b82772e5b70f199e8e53ce557ce0
• Tree: 159eff82ed9fae10ac3627c1df6347bf1790401b
• Head parent: 369804e144232a58dff8c594c991a636425ce9e2
• Integration merge parents: original PR head e948eb8ead6ea5bbeb7f356196f14dd29557c837, current main 6e19f1418718e9e33fdfa2d0d6c03054c5768470

The successor differs from main in exactly two declared paths:

• vinci/test/402-escalation-no-downgrade.mjs
• vinci/test/run.sh

There was no overlap between changes added to main since the original PR base and either PR path, the classifier, or the four consumer extensions.

Checks

• npm run check is clean - errors, warnings and infos alike
• ./test.sh passes (not applicable: this is a Vinci-harness-only test change)
• Added a test for anything that could regress
• Added a ## [Unreleased] entry (not applicable: test-only, no product behavior changed)
• No package-lock.json changes
• No emojis in commits, title, or comments

Focused Node 22 checks passed on the exact successor:

• node vinci/test/402-escalation-no-downgrade.mjs
• node vinci/test/402-classification-integration.mjs
• node vinci/test/no-downgrade-integration.mjs
• all four individual VINCI_TEST_402_CONSUMER selections

Mutation verification: changing the classifier's account result to transient made every selected consumer independently exit 1 with its own Missing expected rejection assertion (advisor-in-flight, council-in-flight, scope-in-flight, and loopbreak-in-flight). Separately replacing each consumer's preserved cause with a new error carrying the same message and status: 402 made every selected consumer independently fail the strict cause-identity assertion. Restoring production sources made the complete and individually selected matrices pass. No production mutation remains.

The supported local Node 22 offline harness passed all groups through 365/365 units, worker suites, containment broker 42/42, package first-party 30/30, loopbreak 84/84, scope 83/83, unattended 96/96, shell 33/33, coding-agent 101/101, Crew 184/184, resultbudget 15/15, and the first UI suite 33/33. The final 23-case UI group stopped emitting progress on this macOS host and was interrupted after 254 seconds; the harness then completed its native PTY check successfully and exited nonzero only for the interrupted UI group. Fresh exact-head hosted Node 22/24 lanes remain the authoritative full-harness gate. An earlier npx --package node wrapper run is not counted because it contaminated nested npm execution; rerunning directly with the cached Node 22 binary made package first-party pass 30/30.

Packaging evidence:

• Network-free build passed.
• Unsigned development package 0.0.51 was created.
• From an unpacked tarball, all 100 relative extension imports resolved; both launcher dispatch targets and 34 imports resolved; vinci worker reached its usage refusal; packaged CLI reported 0.0.51.

Fresh exact-head hosted checks are all green:

• Run 33890967535: check job 101082244752, offline (22) job 101082244616, upstream-suite job 101082244278, sandbox-macos job 101082244680, and offline (24) job 101082244851.
• Run 33890967341: build-check-test job 101082244590.

Required final review: GO for exact head 240ddebe1948b82772e5b70f199e8e53ce557ce0 / tree 159eff82ed9fae10ac3627c1df6347bf1790401b, with no actionable findings. The reviewer independently reproduced the cause-identity mutation failure and the restored positive path.

Anything reviewers should know

The provider stub throws the original error object with status: 402, and the assertion checks that the terminal wrapper retains that error as its cause. Two 402 variants run through all four consumers. A second same-class 402 and a successful cheaper response remain queued as tripwires, proving neither retry nor downgrade occurs.

No billing policy or runtime code changed. Outbox, CFT, provider/model, GPU, and Sentinel boundaries are unchanged.

[thegeorgepu]

GO at exact head e948eb8ead6ea5bbeb7f356196f14dd29557c837.

I verified both real 402 variants through the production advisor, council, scope, and loop-recovery consumers. Each case terminates with the original status: 402 retained as the cause, never attempts a same-class retry or cheaper fallback, and emits no unavailable-class continuation. The queued second 402 plus successful Forte response is a valid tripwire: forcing the classifier to return transient makes the test fail at advisor-in-flight with Missing expected rejection after that retry/fallback path becomes reachable.

The focused test and adjacent classifier/no-downgrade tests pass; npm run check passes; the aggregate harness locally reached and passed the new group, and all hosted aggregate/check jobs are green. The branch is ahead 1 / behind 0 against main.

No actionable findings.

This review is AI-generated by Codex.