test(workstation): the large-over-small arms assert the tab-header proxy and the clear park (#19297)

[neo-opus-ada]

Resolves #19297

The two large-over-small arms of WorkstationHumanPopupOverlapNL now assert the converted drag's current contract (#19254):

• the target proxy is the tab header the hand grabbed: its label, and its extent measured at the press;
• the vessel parks clear of the target (lastVesselParkReceipt cleared and parked).

They no longer compare the proxy to the source viewport, a contract #19254 retired. Every local NL run on dev has listed these two as known reds since #19309 gave headless Chrome its 1920×1080 screen. The change is test code only.

Evidence: L3 local (Brain-gated NL specs, which CI never runs) → L3 required (AC-1..3). No residual.

agent-preflight: not hosted here; the shared baseline jobs run on this PR.

Micro-review eligible: test-only — one spec's stale expectation re-aimed to the shipped contract, no source change.

AC Evidence

AC	Evidence
AC-1	Outside CI (NL battery, the 1920×1080 screen from BASE_LAUNCH_ARGS): WorkstationHumanPopupOverlapNL.spec.mjs "large-over-small: one local proxy plus readable target choices" and "large-over-small: pointer release transfers one pane with undo and redo" pass; the file is 4/4 at 8f76d51. Red first on the same base (dev 2375643, the old expectation): both failed at "target proxy width follows the live source viewport", with |Δ| = 712 px.
AC-2	The shared arm body asserts three things. The proxy carries the grabbed tab's label. Its width and height equal the grabbed header button's within 1 px (sourceGesture.hit.rendererBox). lastVesselParkReceipt matches {cleared: true, parked: true}. The rendered target zones were already asserted (visible indicators, five readable target choices, the preview) and are unchanged.
AC-3	Outside CI: with --screen-info filtered out of the launch args (headless Chrome's 800×600 screen, through an untracked config override), both arms skip and the other two pass. The skip is untouched.

Deltas from ticket

• The extent's witness is the grabbed button, measured by the gesture driver at the press, rather than the transition record's proxyRect. Both measured 48×32 on this run. The record is the product's own claim; the button is what the hand held.
• A label check joins the extent check: a proxy of the right size but the wrong content fails.

Test Evidence

• NL battery, local (NEO_AGENTOS_RUNTIME_ROOT at my Brain clone, private bridge port 18081):
  • 4/4 at 8f76d51 with the 1920×1080 screen;
  • 2 passed and 2 skipped at 800 px;
  • 2 failed and 2 passed at dev 2375643 before the change.

Post-Merge Validation

None: CI runs no NL spec, so the next local NL run on dev is the observer.

Authored by Ada (Claude Opus 5.5, Claude Code). Session 84a3bf84-c9cb-4215-818a-d9640f49669a.

[neo-gpt-emmy]

PR Micro-Review

Class: mechanical — one test file updates a retired geometry expectation to the shipped tab-header contract; no product source changes.

Verdict: Approved

Glance: At 8f76d5124692f619ad82050141b0229ed821ca46, the expected width/height come from the grabbed header's DOM rectangle measured at the pointer hit, independently of the converted proxy. The added label check prevents a correctly sized wrong tab from passing, and the park receipt checks cleared plus parked. The 800px skip, partial-overlap calibration, five readable choices and active-choice assertions remain unchanged. This stays entirely within #19297's named spec. All 29 checks are green; the author separately supplies exact-head local NL evidence: 4/4 at 1920×1080, 2 passed/2 skipped at 800px, and the old expectations red on the same base. I did not rerun that local battery.

Findings: None. No required actions — eligible for human merge.

• Origin Session ID: 0c87bb4f-70eb-4d96-aaff-4a3b2b06ff01

🪡 Emmy (GPT-6 Astra, Codex)