fix: return full Perplexity responses and reliable completion status

[MrJ55]

Fix response extraction truncation + stuck WORKING status

Problem

1. Truncated answers: getAgentStatus collected every [class*="prose"] block but returned only the last one, so long Perplexity answers came back as their final fragment (e.g. only the last bullet of a multi-section answer).
2. Stuck WORKING: the completion heuristic matched streaming words ("Working", "Searching", "Analyzing"…) against the whole page body — and the answer text itself often contains those words — so comet_poll never flipped to completed once the answer was done.

Changes (src/comet-ai.ts only)

• Join all prose blocks with containment dedup (Perplexity nests prose elements; innerText includes descendants) → complete, non-duplicated answers.
• Raise the response cap 8000 → 30000 chars.
• Reorder status detection: hasAskFollowUp && hasProseContent && !hasActiveStopButton → completed before the working-text branch. ("Ask a follow-up" UI only appears when streaming finished.)
• Harden against undefined CDP evaluate results (dead tab / navigate race / closed browser) — no more Cannot read properties of undefined (reading 'toUpperCase').

Verification

• Tested end-to-end against live Perplexity threads (search mode) via the MCP gateway: comet_poll returns the full deduplicated answer; status flips to completed even when the answer text contains streaming words.
• Manual CDP evaluations confirm status=completed with the complete response length.

Notes

• dist/ is gitignored; only the TypeScript source is included.