feat(extractor): filter redundant ghost user_ids during cc_session sync

[hrdAI3]

Summary

Skip "redundant ghost" cc-session uploads from the Matrix-Riven collector so the workboard doesn't double-count the same physical CC session under two user_ids.

Background

Matrix-Riven's realtime path falls back to \${unix_user}@\${hostname} when git config user.email is unset (and pre-PR #3 over there, also when the digital-twin config had a real identity — that PR fixes new sessions but historical data + clients that haven't reinstalled yet still produce ghosts). On the collector this surfaces as same-machine pairs:

• Stop hook → transcript → tagged hrdai@qq.com (real)
• Realtime hooks → cc-status → tagged 19723@hut (ghost)

After cleanup the live /api/users list still has 4 sole-identity ghosts whose data only exists under the hostname form:

lv@lvjiawendeMacBook-Air.local              → 吕嘉文 (12MB+ transcripts here)
blink@BlinkdeMacBook-Air.local              → 孙润峰
zhangziyi@zhangziyideMacBook-Air-2.local    → 张紫翊 (16 transcripts here)
alexpeng@pengchengdeMacBook-Air.local       → 彭丞

These four MUST NOT be filtered — they ARE those people's only upload identity until they set git config user.email and reinstall the client.

What this does

Two pure helpers in src/lib/ghost_user.ts:

• isGhostUserId(userId) — shape predicate. Real-email TLD allowlist + .local hostname suffix + bare-hostname (no dot) classification. Conservative on unknown TLDs (leans real).
• canonicalSessionId(rawId) — strips the trailing .cc-status the collector appends to snapshot listing entries.

syncCcSessions gets a pre-pass:

1. List sessions for every non-ghost user inside the lookback window.
2. Collect transcript session_ids (the entries WITHOUT .cc-status suffix) into a global Set.

Main loop then for each ghost user's session: if canonicalSessionId(sess.id) is in that Set → skip ingestion (don't fetch raw, don't emit events, don't write to events.jsonl), bump a new summary.redundantGhostsSkipped counter, advance lastMtime so we don't re-evaluate the same entry next sync.

Sole-identity ghosts (their session_ids have no non-ghost peer) pass through untouched and emit cc.* events under their hostname-form user_id, same as before.

Tests

8 cases in tests/lib/ghost_user.test.ts — real email TLDs (qq.com / outlook.com / 163.com / nb-ai.com / gmail.com / GitHub users.noreply.github.com), bare hostname (19723@hut), .local mDNS (blink@..., lv@..., zhangziyi@..., alexpeng@...), edge cases (empty, no-at, trailing-at), and the unknown-TLD conservative-lean (person@company.example → not flagged).

All 8 pass.

bun run test shows 5 pre-existing failures (bootstrap/extract.parseOrgChart, evolution/diff.path-allowlist, lib/agents.writeProfile+updateState) — confirmed unrelated by stash-and-rerun on clean main.

Not done in this PR

• Backfilling the existing redundant cc-status events already in events.jsonl from earlier syncs. Pre-existing data is left alone; the filter only affects future syncs. If the events.jsonl is large enough that the historical ghost noise matters, that's a separate one-shot script.
• Mapping ghost user_id → real Chinese name. Out of scope here; lives in the team's identity registry if/when added.

Test plan

• bun run typecheck — no new errors in changed files (pre-existing errors in vector_match + workboard are unrelated)
• bun run test tests/lib/ghost_user.test.ts — 8/8 pass
• Stash-and-rerun confirms unrelated test failures are pre-existing
• Manual: next live sync run should report redundantGhostsSkipped > 0 if any ghost-paired sessions land in window

🤖 Generated with Claude Code