Add CI contract tests for the ATM Send-To picker protocol

[randlee]

Wyvern CI contract-test request

Request

Please add a Wyvern CI contract test for the ATM Send-To picker. The test
should execute the released picker with the shared fixtures and fail on any
wire-contract drift. This issue deliberately contains the complete contract
so the Wyvern implementation does not need to depend on atm-core source.

PickerInput (schema version 1)

The picker receives exactly one JSON object on stdin:

{"schema_version":1,"teams":[{"id":"…","name":"…","members":[{"id":"…","name":"…","host":"…","cwd":"…","status":"active|idle|dead"}]}]}

schema_version is the compatibility gate. The page must explicitly declare
the version it implements and reject an unknown input version rather than
guessing at field meanings. Additive contract evolution increments the
version; it must not silently widen version 1.

PickerOutput (schema version 1)

On confirmation, the picker writes exactly one JSON object to stdout:

{"schema_version":1,"recipients":["member-id","…"],"note":"optional one-liner"}

Version 1 has exactly those three keys. recipients is a non-empty
multi-select list of member IDs and note is optional free text. The shared
fixture corpus in atm-core is:

• docs/plans/phase-aq/fixtures/picker-input-v1.json
• docs/plans/phase-aq/fixtures/picker-output-v1.json
• docs/plans/phase-aq/fixtures/picker-output-unknown-schema.json

The CI test should consume the same bytes (or a synchronized copy checked by
hash) rather than maintaining a subtly different schema fixture.

Process contract

• stdin contains one PickerInput object and nothing else;
• stdout contains one PickerOutput object and nothing else — no logs, HTML,
  progress, or diagnostics;
• diagnostics belong on stderr;
• cancel exits nonzero and emits no stdout at all;
• --version is parseable as MAJOR.MINOR.PATCH, returns quickly, and is
  suitable for a bounded compatibility probe;
• launch-to-interactive is approximately 1 second or less; record the
  measurement in CI and fail or flag a regression according to the release
  policy.

The test should cover confirmation, cancel, unknown schema rejection, exact
stdout shape, stderr-only diagnostics, the version probe, and the cold-start
budget. Please link the resulting workflow job or PR here when implemented.

[randlee]

Source of truth by reference (supersedes the inline sketch above where they differ)

Per Rand: the precise schema requirements should land in Wyvern as integration tests that reference atm-core's canonical files, so drift is caught on both sides. Pin to an atm-core ref and hash-check the fetched fixtures; do not maintain a divergent copy.

atm-core (branch feature/aq-5-surface-evidence, ref dcff4bac3; will move to integrate/phase-aq → develop at Phase AQ closeout):

What	Path in atm-core
Contract doc (authoritative, incl. the invocation-shape correction)	docs/plans/phase-aq/fixtures/wyvern-pick-member-contract.md
PickerInput v1 canonical bytes	docs/plans/phase-aq/fixtures/picker-input-v1.json
PickerOutput v1 canonical bytes	docs/plans/phase-aq/fixtures/picker-output-v1.json
Unknown-schema rejection fixture	docs/plans/phase-aq/fixtures/picker-output-unknown-schema.json
Adapter (macOS/Linux) — generates wizard.json, invokes Wyvern, unwraps WizardResult.data	scripts/send-to/atm-send-to.sh
Adapter (Windows)	scripts/send-to/atm-send-to.ps1
Bounded version/contract probe (1.5 s deadline)	scripts/send-to/probe_wyvern.py
Six-case degradation harness (absent / below-pin / unparsable / hang / missing-asset / unknown-schema)	scripts/phase-aq/run_aq5_wyvern_degradation_evidence.py
Adapter + degradation tests (fake-wyvern fixtures emit the real WizardResult envelope)	.just/tests/test_send_to_surface.py, .just/tests/test_picker_exclusion.py
Pinned Wyvern version	WYVERN_PIN (single source; AQ6 preflight bumps/pins back from it)

Corrected process contract (real Wyvern, verified against v0.5.0 while building #140)

• There is no wyvern --picker <page> flag. The picker runs as a wizard: the adapter writes wizard.json with config = PickerInput and runs it; Wyvern's terminal stdout is the full WizardResult envelope {"button":"finish","data":<PickerOutput>,"stack":[...]}; the adapter reads .data.
• Cancel = non-finish button / nonzero exit with no PickerOutput; diagnostics on stderr only; --version parseable MAJOR.MINOR.PATCH within the 1.5 s probe.
• idle/dead members must be rendered disabled (non-routable), not merely styled; unknown schema_version is rejected, never guessed.

Requested Wyvern CI shape

1. Keep feat(examples): atm-pick-member wizard contract page (atm-core Send-To, issue #139) #140's Playwright L2 test (tests/l2/wizard-atm-pick-member.spec.ts) as the page-level contract test.
2. Add an integration job that checks out atm-core at the pinned ref, runs scripts/phase-aq/run_aq5_wyvern_degradation_evidence.py and .just/tests/test_send_to_surface.py against the real Wyvern build on PATH, and fails on any wire drift.
3. When the contract evolves, bump schema_version and the atm-core ref together; the hash check makes silent widening impossible.

[randlee]

Confirming the final atm-core paths and head after the real-invocation fix round.

atm-core branch feature/aq-5-surface-evidence, head 8cb881dfe99879db3cc09bf089757bc947cbb523 (moved from dcff4bac3 in the by-reference table above — that commit only had the fixture/exclusion fixes; this head adds the actual wizard.json-in/WizardResult.data-out wiring in atm-send-to.sh/.ps1). Will move to integrate/phase-aq → develop at Phase AQ closeout as noted.

All paths in the table above are correct and present at this head. Two small corrections/clarifications:

• WYVERN_PIN is not a single shared source today — it's the literal "0.5.0" duplicated in both scripts/send-to/atm-send-to.sh (WYVERN_PIN=) and scripts/send-to/atm-send-to.ps1 ($wyvernPin =), kept in sync by convention/AQ6 preflight, not by one constant both scripts import. Worth knowing if a Wyvern-side integration job wants to read "the" pin — it should read from atm-send-to.sh (the canonical shell adapter) unless both are checked for agreement.
• .just/tests/test_picker_exclusion.py is not a Wyvern-degradation test file — it's the R4 dead/idle-exclusion contract test (all four picker adapters, hermetic). The Wyvern-degradation-specific coverage lives in .just/tests/test_send_to_surface.py::test_wyvern_degradation_cases_fall_back_and_still_send (six-case fallback) and the new ::test_generated_wizard_json_matches_contract_shape (asserts the generated wizard.json matches the contract doc's shape). Both now use fake-wyvern fixtures that emit the real WizardResult envelope with .data, not the old bare PickerOutput.

Also new at this head, not yet in the table:

What	Path in atm-core
Vendored pick-member.html asset (byte-identical copy of randlee/wyvern#140's page, used as the --ui-root asset)	scripts/send-to/pick-member.html
Local end-to-end transcript: real wyvern#140 build driven by atm-send-to.sh unmodified (exit 0)	docs/plans/phase-aq/evidence/AQ5/wyvern-real-invocation-local.md

That transcript was produced by building randlee/wyvern#140 at commit 958b5102e977f30f812213d5ae08c1420828bead locally, then running scripts/send-to/atm-send-to.sh end to end against the real binary (headless via Wyvern's own WYVERN_VIEWER=none/WYVERN_DIALOG_URL_FILE contract, driven over /api/wizard/state + /api/wizard/finish, the same HTTP surface a real browser click uses) — confirming the config-in/.data-out shape round-trips through a real Wyvern process, not just the hermetic stub fixtures.

Corrected process contract and requested Wyvern CI shape above both match what's implemented; no changes needed there. The same table (with this head's corrections folded in) is going into #140's README now.