feat: conversation fabric types, P0 CDP spike, and Grok provider discovery

[MrJ55]

Summary

Adds the conversation-fabric type contracts, the P0 CDP-concurrency findings, and the P2 Grok provider discovery to comet-mcp — the first concrete steps toward the multi-provider backbone described in docs/build-plan.md and ADR 0001. No existing behavior changes; all new files.

What's included

P1 — Conversation fabric type contracts (ADR 0002, src/types/)

• conversation.ts: ConversationEnvelope (distinct idempotency key / correlation ID), DeliveryReceipt with 7 statuses (queued|sent|accepted|completed|blocked|timed_out|unknown — unknown is never silently resent, per ADR 0001 §5), append-only ConversationEvent log, Provenance (attribution is an audit control, explicitly not a safety claim), content persistence modes (full|redacted|none), conservative relay defaults (approval-required, unapproved).
• provider.ts: normalized ProviderState union, PollResult (truthful hasStopButton; extraction-provenance flags preserve the Bug WSL Support: Add COMET_HOST and COMET_PORT environment variables #1/Bug Stop closing user's existing browser tabs on connect #2 fix lineage), TabSession (P3 reconnect/dedup anchors typed now), HealthReport (per-control hook-resolution source for P8 drift detection), ChatDriver adapter contract (open/ask/poll/stop/reset/health).

P1/P2 — Provider registry entries (src/providers/)

• grok.ts: HIGH-confidence entry from live discovery (selectors, heuristics, state machine). Documents the verified finding that Grok's Fast model never renders a stop button — streaming state must use the "Working for Xs" indicator.
• perplexity.ts: MEDIUM-confidence entry extracted from the existing comet-ai.ts contract + session-handoff verification (the P1 refactor seed).

P0 — CDP concurrency spike (test/integration/cdp-concurrency-spike.mjs, docs/p0-cdp-concurrency-findings.md)

• Zero-dependency spike: 2→3→5 concurrent tabs, one CDP session per target, 60s per phase. Result: 0 failures, 0 disconnects, 0 cross-tab effects; all 5 sessions stable; latency flat (p50 ~2ms). Gate passed at the ADR 0001 target of 5 sessions. Recommended default cap: max_concurrent_tabs: 5, configurable.

P2 — Grok discovery workflow (test/integration/grok-discover.mjs, test/fixtures/grok/, docs/runbooks/grok-provider-discovery.md)

• Offline/on-demand selector miner (not in the hot path) with live PONG validation, state fixtures (idle/typing/streaming/completed), and a --diff mode that re-runs discovery and diffs the provider entry against the previous run (workflow step 9).

Docs

• docs/build-plan.md, ADR 0001 + 0002, design syntheses (00–02), reference docs, runbook.

Notes for reviewers

• tsc --noEmit passes with both provider entries consuming the shared types.
• The type layer is declarations only — the event-log store, concrete ChatDriver implementations, and the Perplexity refactor are next (P1 runtime work).
• Fixture HTML files are sanitized snapshots (scripts/styles/svg stripped); they capture structure for synthetic tests.
• test/integration/out/ run artifacts are gitignored as regenerable.
• Verification: PONG round-trip against live grok.com in the authenticated Comet profile (Chrome/150.0.7871.230), three consecutive discovery runs with a clean diff.

Related

• ADR 0001: Browser-tab transport and relay defaults
• ADR 0002: Conversation fabric type contracts
• Build plan: docs/build-plan.md