fix(task-board): duplicate check survives a reasoning model's answer

[viktormarinho]

Summary

Second production follow-up to #7365 (#7376 made the outcome visible; this fixes what it revealed).

With duplicateCheck in the output, the throwaway-org test now showed skipped: unparseable answer from qwen/qwen3.5-flash-02-23 on two of three calls. The org's fast tier resolves to a reasoning model: it thinks before it answers, so the JSON arrives after a <think> block, after narration that may itself contain braces, or not at all within the 1200-token budget.

Changes

• parseModelJson strips <think>…</think>, ignores code fences anywhere, and tries every top-level balanced {…} from the last one backwards, so narration before the answer never wins.
• maxOutputTokens 1200 → 8000, room for the model to reason before the JSON.
• The skip reason now carries finishReason and the answer length, so a truncation reads as finish=length instead of a generic "unparseable". The warn log gets the tail of the answer rather than the head.
• Two parser tests: a <think> block with braces inside, and narration with an earlier JSON object before the real one.

Testing

28 tests in duplicate-check.test.ts, bun run check, bun run lint, bun run knip, bun run fmt. Will re-run the production scenario on the same org once deployed and report on this PR.

---

Summary by cubic

Fixes task-board duplicate check for reasoning models so their JSON answer is used instead of being skipped as unparseable.

Bug Fixes

• parseModelJson strips thinking blocks, ignores code fences, and tries top-level JSON objects from last to first.
• Raises the fast tier maxOutputTokens from 1200 to 8000, which may increase cost/latency when the extra budget is used.
• Skip reasons now include the finish reason and answer length so truncation is distinguishable from malformed output.
• Adds parser tests for a reasoning block with braces and narration containing an earlier JSON object.

^{Written for commit b5e621d. Summary will update on new commits.}

[viktormarinho]

Production data behind this PR, from the throwaway org:

• With the org's default fast pick (a qwen flash reasoning model): 3/3 skipped: unparseable answer once the board had a dozen cards. Longer boards mean longer thinking, so the JSON never fits in 1200 tokens or arrives after a <think> block.
• With the org's fast tier pinned to anthropic/claude-haiku-4.5 via ORGANIZATION_SETTINGS_UPDATE, on the code already deployed (fix(task-board): duplicate check answers via generateText and reports its outcome #7376, before this PR): 3/3 matched on paraphrases (EN↔PT, both directions), 2/2 no_match on related-but-distinct tasks (same area, different change).

So the judgement and the confidence gate are right; the remaining failure is purely the answer format of reasoning models on the default pick, which is what this PR addresses. Orgs that pin a non-reasoning fast model already work without it.