[ffroliva]

Summary

Live verification overturned my own fix from an hour earlier. That is the headline, and it is the reason this PR exists.

#739 shipped a guard raising FlowHostMigratedError, asserting flow.google.com "renders no '+ New project' control gflow can drive". A $0 run of _enter_editor with no project id created a project there in one click.

The claim was an unproven negative — generalised from a sweep of two other surfaces (the project composer and the character editor) to a third that was never probed. The gallery is where NEW_PROJECT_SELECTORS actually lives, and I never measured it.

Refs #730

What the gallery actually shows

denon82, control 47 ligature nodes, $0:

add* ligatures            {"mat-icon|add": 1}
[0] add_2 / i.gs          0
[1] add_2 / i             0
[2] add_2 / role=button   0
[3] text-matches regex    ERR Error
[4] has-text('New project')  1     <- the ONLY match

The ligature finding was right — add, never add_2. The conclusion was wrong. The CTA is present and clickable; every structural Tier-1 entry missed, and it survived purely on the English text fallback — the anti-pattern Tier 1 exists to prevent, and one that fails outright on a non-EN migrated profile. My guard would then have told that user the control does not exist: false, and a dead end.

Fixed with a class-only add anchor leading Tier 1, covering both carriers. Re-measured live: entry [0] now matches 1 where every previous entry matched 0.

Also deleted: a selector that never worked

button:text-matches('^\+\s+\S+$', 'i') is not a valid Playwright selector. It raises on every evaluation — observed twice against the live gallery — and except Exception: continue swallowed it. It has never matched anything on any host, while costing a round trip per attempt.

Two guard tests caught this change, and one of them was the problem

• test_new_project_selectors_icon_leads pinned the add_2 selector exactly. Updated to the measured anchor, with the measurement as its stated reason.
• test_new_project_selectors_plus_regex_is_anchored asserted the regex was anchored — a property of a selector that never ran. It inspected the string instead of evaluating it, so it passed happily for as long as the selector was broken.

Replaced with the guard that would have caught the real bug: every shipped cascade entry must parse against a real CSS engine (.count() on a blank page — no account, no network, no credits), plus a companion asserting the known-bad selector does raise, so the guard cannot quietly stop discriminating.

Validation

• pytest -q --cov — 4069 passed, 7 skipped
• pyright src 0 errors · ruff check / format --check clean
• doc links · hygiene · website mirror · memory round trip · PII scan
• Live re-measurement of the add anchor on the migrated gallery

E2E: the CTA path is only reached when a generation runs without --project. The live gallery probe above is the direct evidence for the changed selector, and it is stronger than an e2e for this purpose: it counts what each cascade entry matches, rather than confirming that something in the cascade worked. Not re-running the character-create e2e — that path passes --project and never touches this code.