feat(build): ground "I'll wire that"/"it's in the queue" in real Gitea issues (#774, Gap 2)

[urbantech]

Summary

Closes Gap 2 of #774.

Real incident: a paid, git-provisioned founder said "The stripe integration" (a topic phrase, not an imperative edit — detectEditIntent correctly stayed conservative and never fired). Cody still replied "I'll have this wired in the next run" — gateInstructions unconditionally scripted that phrasing for every change/feature message, with zero task or issue ever actually created. fetchBacklogSummary's own "backlog" was purely synthetic (derived from selectPrimitives), with no connection to lib/git/gitea-client.ts at all — Cody had no real system of record to cite either.

Fixed with two combined detection layers feeding one real action:

1. Cheap, deterministic heuristic (no LLM call): auto-qualifies any short (≤6 word), non-question message that reaches here without matching detectEditIntent's imperative-verb pattern — the exact "The stripe integration" shape from the real incident.
2. Secondary, single-purpose classifier call (completeText, bounded to 6s, fails closed to "no"): handles longer/ambiguous phrasing the heuristic doesn't resolve.

Either path files a real Gitea issue in the company's own repo (new listIssues + the already-proven createIssue) before the system prompt is built. gateInstructions' "I'll wire that"/"it's in the queue" framing is now conditioned on real work (an edit dispatch OR a filed issue) having actually happened this turn — never said as scripted filler otherwise.

fetchBacklogSummary also now surfaces the company's real open/recently-closed Gitea issues alongside the synthetic primitive-derived list, so "it's queued"/"already built" claims are grounded in the same system of record the nightly loop and the SMS-to-issue path (#744) already use.

Test plan

• npx tsc --noEmit — clean
• npx vitest run — full suite: 396 files, 5615 tests passed, 0 failures
• New gitea-client-listissues-774.test.ts (9 tests): success/failure/thrown-error paths, state/limit params, unconfigured degradation
• New ask-real-work-grounding-774.test.ts (13 tests): heuristic path files an issue with no classifier call, plain questions never file, classifier path both ways, timeout/failure degrades safely, no double-filing when an edit was already dispatched, free-tier/unprovisioned never file, system-prompt framing conditioned correctly both ways, real Gitea issues surfaced in the backlog block, graceful degradation when Gitea is unreachable
• Confirmed zero regression across all existing ask/edit-intent test suites (71 tests)