feat(mcp): MCP v2 — deep_dive tool, fit fix, schema honesty, GITHUB_TOKEN + timeout

[New1Direction]

Summary

Turns the MCP proof (#33) into something real: a third tool, two correctness fixes the survey caught, and the rate-limit + timeout hardening it needed. All extension modules are reused verbatim; only mcp/ changes plus a backward-compatible optional param on two shared fetch helpers.

What's in it

New tool — deep_dive({ repo }) — plain-English explanation of how a repo works, with gaps, assumptions, self-test questions, and per-claim confidence, plus the underlying atoms + lineage. Reuses deepdive.js verbatim (atoms → lineage → Feynman, three model calls — the heaviest tool).

Bug fix — scan_repo fit was always undefined. The schema declared fit required, but parseClaudeResponse never produces it (the extension derives it separately via verdict.js deriveFit). Now a pure buildScanResult calls deriveFit, the schema declares the real { level, label, why } object, and bottom_line is surfaced too.

Bug fix — blueprint_scene outputSchema lied. It advertised edges as { source, target, label } and flat nodes, but buildBlueprintScene returns engine-shaped edges { id, from, to, rel, note, userDrawn } and nodes with ref/layer/pinned. Any client trusting the declared schema would mis-parse. Schema now matches the real scene.

GITHUB_TOKEN pass-through — optional opts.githubToken threaded into fetchGitHub (fetcher.js) and fetchSource/ghJson (deepdive.js). No token → byte-for-byte the old anonymous path, so the extension is untouched. With one, GitHub's 60 req/hr → 5000/hr (blueprint/deep_dive make 10+ calls each and would otherwise 403 mid-scan).

Hard timeout on the Anthropic call — AbortController + ANTHROPIC_TIMEOUT_MS (default 60s) so a stalled connection can't hang a tool call forever (mirrors the extension's per-provider timeout).

Reuse / invariants

• Extension stays zero-dep; all deps remain isolated under mcp/.
• Shared-module changes (fetcher.js, deepdive.js) are purely additive optional params — existing extension callers pass nothing and behave identically (confirmed: the extension's own 849 tests still pass).

Test plan

• npx vitest run → 857 passed (was 849; +8 new offline MCP tests, no key/network)
  • ghOpts env plumbing; buildScanResult fit derivation (strong + risky); buildDeepDiveResult assembly
  • full runScanRepo / runDeepDive via mocked global.fetch — asserts the Authorization: Bearer header appears only when GITHUB_TOKEN is set, and deep_dive makes exactly three model calls
• Runtime ESM import check — all three tools load; blueprint edge schema uses {from,to}
• npx eslint . → 0 errors (25 warnings, baseline)
• node --check on all touched files

Not in this PR (follow-ups)

• npm / PyPI / GitLab inputs for scan_repo (the fetcher already supports them — only repo-input.js is GitHub-only)
• Multi-provider (reuse the extension's providers.js registry)
• A tools/list structural smoke test

⚠️ Not live-tested against a real key/network (no key in CI) — exercised entirely through mocks.