fix(global-discover): bucket codex by originator + read 128KB for CC cwd

[0xDevNinja]

Summary

Two patches from #1315 in one diff.

1. Codex session bucketing. scanCodex now normalizes payload.originator into { desktop, exec, claude_code, other } and surfaces the breakdown at tools.codex.originators and per-repo codex_originators. Existing codex totals stay (additive — no consumer break).
2. CC undercount. extractCwdFromJsonl reads 128KB instead of 8KB. Recent Claude Code / CCR JSONL files often open with a queue-operation event 30-50KB long that has no cwd — the old 8KB read truncated the line, JSON.parse failed, and the whole project dir was silently dropped. Same buffer size scanCodex already uses.

Fixes #1315.

Why

/retro global narrated "codex was the primary execution tool, 414 sessions across 7 repos" when codex actually drove dev for one repo's middle phase. The other ~309 codex_exec entries were CC firing codex as cross-model review subagent. A single bucket can't tell those apart.

For the CC count: @Akagilnc traced ~450 missing files in one repo's 31d window to the 8KB cap (issue thread). First-line queue-operation events are 30-50KB on recent CC versions; the parser never reached the later events that carry cwd.

Shape

// tools.codex now:
{
  "total_sessions": 414,
  "repos": 7,
  "originators": { "desktop": 92, "exec": 309, "claude_code": 13, "other": 0 }
}

// per repo:
{
  "name": "ak-ai-vela",
  "sessions": { "claude_code": 12, "codex": 98, "gemini": 0 },
  "codex_originators": { "desktop": 1, "exec": 97, "claude_code": 0, "other": 0 }
}

Summary format inline shows Codex:98 (desktop=1, exec=97, cc=0) per repo + a top-line Codex originators: ... rollup.

Originator normalization (case-insensitive, matches values observed in ~/.codex/sessions/):

• "Codex Desktop" / "codex_desktop" → desktop
• "codex_exec" / "codex exec" → exec
• "Claude Code" / "claude_code" → claude_code
• anything else → other (not dropped — future originators land here visibly until we map them)

Also adds a CLAUDE_PROJECTS_DIR env override on scanClaudeCode so the CC regression test can stage a fake project dir; mirrors the existing CODEX_SESSIONS_DIR knob.

Out of scope

Problem 3 from the issue (annotate /retro global that "sessions = tool invocations / file count", not interactive dev) is a narrative-side change in the retro skill template. Reasonable to file separately.

Tests

7 new in test/global-discover.test.ts:

• 'Codex Desktop' originator → desktop bucket.
• 'codex_exec' → exec bucket.
• 'Claude Code' → claude_code bucket.
• Unknown originator string → other (verifies nothing is dropped).
• Per-repo codex_originators sums to per-repo sessions.codex.
• tools.codex.originators shape + total parity with tools.codex.total_sessions.
• CC JSONL whose first line is a >30KB queue-operation event still resolves cwd from a later event.

bun test test/global-discover.test.ts
# 26 pass, 0 fail (19 existing + 7 new)