fix(e2e): drive Claude model selector with trusted clicks + assert-closed (#501)

[rosscado]

Why

On a 2026-07-05 host sweep, selectClaudeModel() left claude.ai in a stuck state: a full-page gray overlay with an empty floating-menu remnant persisted from model selection through the rest of the run, blocking page.click("#saypi-callButton") ("could not click call button") and recording transcript=false reply=false for a run where SayPi itself was never exercised. Immediately re-running with --claude-model=keep (which skips this function) produced a fully clean run — pinning the fault to this step, not the extension.

Root cause

selectClaudeModel() opened the menu and selected the radio with synthetic page.evaluate(() => el.click()) clicks, then closed with keyboard.press("Escape"). Synthetic clicks dispatch without the microtask checkpoints a trusted click gets, which desyncs Claude's Radix-style dismissable-layer open/close bookkeeping — the menu content closes but its full-page backdrop is stranded over everything. Same class of bug as the #494 voice-menu race. The distinguishing factor vs selectSaypiVoice (unaffected) is the close path: model used a bare Escape; voice re-clicks its toggle.

Fix

• Open + select with Playwright's trusted locator.click() (getByRole('menuitemradio').filter({ hasText }).filter({ hasNotText: /currently unavailable/i })), never synthetic evaluate(el.click()).
• Replace the bare Escape with an explicit assert-closed guard: waitForFunction until no [role='menuitemradio'] remain, with a trusted Escape fallback if the menu lingers.

The page's interactivity is now a verified post-condition of the function, satisfying the issue's acceptance criterion.

Scope / tests

Test-harness script only (scripts/e2e-host-sweep.mjs); no product code. node --check parses; full merge gate green locally (204 files). True verification is a real claude sweep run — see PR comments for result.

Fixes #501

🤖 Generated with Claude Code

[rosscado]

Real-host verification (Layer 4 CDP, e2e-host-sweep claude with default --claude-model=haiku, so selectClaudeModel runs):

claude: model → Model: Haiku 4.5 Extended (pref: haiku)
claude: decorated=true cf=false transcript=true reply=true auth=true voice=Say, Pi (selected:Jarnathan) model=Model: Haiku 4.5 Extended | saypiErr=0 saypiWarn=0 pageErr=0 netFail=0

Model selection succeeded and the turn completed end-to-end: transcript=true reply=true, no could not click call button note, no stranded overlay, 0 errors. This is the same outcome the issue's --claude-model=keep run produced — now achieved through the model-selection path, satisfying the acceptance criterion.

[rosscado]

Independent reviewer verdict (subagent): APPROVE — the trusted-click change correctly addresses the #501 root cause with no regressions; the API usage (v1.60.0), the null/graceful-degradation contract, and the picked label extraction are all confirmed equivalent to the original.

The reviewer raised one non-blocking finding: the assert-closed guard only checked [role='menuitemradio'] count (menu content gone), which can't distinguish a clean close from a stranded-backdrop-but-content-removed page — the exact #501 failure mode.

Addressed in f7b1125: the guard now requires both content-gone and the trigger reporting collapsed (aria-expanded !== "true"). In the stranded-layer failure state the trigger still reads aria-expanded="true", so this pair rules out the false-pass the reviewer flagged. Comment corrected to describe what's actually verified.

Re-verified on the real host with the strengthened guard (e2e-host-sweep claude, default --claude-model=haiku):

claude: model → Model: Haiku 4.5 Extended (pref: haiku)
claude: decorated=true cf=false transcript=true reply=true auth=true voice=Say, Pi (selected:Jarnathan) model=Model: Haiku 4.5 Extended | saypiErr=0 saypiWarn=0 pageErr=0 netFail=0

The stricter aria-expanded check flips cleanly (no false-negative, no added latency) and the full turn completes. Merging once CI on f7b1125 is green.