fix(hooks): SuccessClaimGate no longer misclassifies shell/CLI liveness as UI liveness

[Luo0oo]

Problem

hooks/SuccessClaimGate.hook.ts fires its WEB-UI VERIFICATION GAP block-back on any Stop-chain response that says a shell hook, CLI tool, or backend subsystem is now live / works end-to-end — even when no browser surface is claimed. The gate correctly closes the incident it was built for (curl-only "admin page is live and locked down"), but the classifier collapses two orthogonal notions (liveness vs. UI-ness) and over-triggers.

Reproduction — this trips the gate with no page, no UI, no HTTP surface:

Memory subsystem now live end-to-end. Both hooks wired via symlink.

Recursive dimension: a message reporting this bug — where the trigger vocabulary has to be quoted for reference — trips the same gate. The quoted-attribution guard in stripNarration covers CLAIM_PATTERNS vocab (shipped|deployed|landed|…) but not LIVE_UI_CLAIM vocab (is live, locked down, works end-to-end), and code-delimited mentions (backticks — the standard example-quoting convention) aren't stripped at all.

Root cause & fix (3 changes, +4/-1)

1. isUi OR-branch collapse. LIVE_UI_CLAIM proves liveness, not UI-ness, but it was an OR-branch of isUi, so every liveness claim self-classified as a UI claim regardless of noun:

- const isUi = WEB_UI_NOUN.test(s) || RENDER_VERB.test(s) || LIVE_UI_CLAIM.test(s);
+ const isUi = WEB_UI_NOUN.test(s) || RENDER_VERB.test(s);

Real UI signal comes from WEB_UI_NOUN (page/site/admin/dashboard/UI/…) or RENDER_VERB (renders/loads/in the browser/…). LIVE_UI_CLAIM stays in assertsLive where it belongs — the gate still requires assertsLive && isUi, so a genuine UI claim (which always names a UI noun or a render verb) still blocks.

2. Mentions vs. assertions. Quoted or code-delimited trigger vocabulary is a mention, not an assertion. Strip markdown code delimiters before pattern-testing, and extend the quoted guard to cover WEB-UI vocab:

 function stripNarration(message: string): string {
   let out = message;
+  out = out.replace(/```[\s\S]*?```/g, " "); // fenced code block — mention, not assertion
+  out = out.replace(/`[^`\n]+`/g, " ");       // inline code — mention, not assertion
   for (const guard of NARRATION_GUARDS) out = out.replace(guard, " ");

   /"[^"]*\b(shipped|deployed|landed|verified|finished|pushed|merged|complete[d]?)\b[^"]*"/gim, // quoted/attributed
+  /"[^"]*\b((is|it'?s|site'?s|page'?s|now)\s+live|went\s+live|locked\s+down|works\s+(now|in\s+the\s+browser|end[\s-]?to[\s-]?end))\b[^"]*"/gim, // quoted WEB-UI vocab — mention, not assertion

The new guard reuses the exact shape of the existing accepted quoted-attribution guard ("[^"]*…[^"]*"), so it carries no new backtracking risk.

Regression coverage

Case	Message	Before	After
A	Memory subsystem now live end-to-end. Both hooks wired via symlink.	block	pass (false positive fixed)
B	bug report quoting now live / locked down in backticks	block	pass (false positive fixed)
C	The admin page is live and locked down.	block	block (2026 incident coverage preserved — admin/page match WEB_UI_NOUN)
D	The dashboard renders correctly and is now live.	block	block (dashboard matches WEB_UI_NOUN)

No new deps, no settings changes — purely a classifier tightening. The repo ships no hook test harness, so there is no test file to update; the four cases above are the manual verification (also confirmed against a deployed copy on WSL2/Bun).

Reproduced against v6.0.0; the affected lines are unchanged on main (v6.0.3).