fix(discover): parse Codex sessions with large session_meta

[sensdiego]

Summary

• Increase read buffer from 4KB to 128KB in scanCodex() — Codex CLI v0.117+ embeds the full GPT-5 system prompt (~15KB) in the session_meta line of rollout-*.jsonl files. The 4KB buffer truncated this line, causing JSON.parse to fail silently for every Codex session. All three originator types are affected: codex_exec (CLI), Codex Desktop (VSCode-based app), and codex_cli_rs (Rust CLI).
• Add CODEX_SESSIONS_DIR env var for test isolation (same pattern as other scanners that accept path overrides).
• 2 new tests: an integration test that creates a fixture with a 20KB session_meta line and runs discovery via CLI, and a unit test that proves the 4KB buffer fails while the 128KB buffer succeeds.

Root cause

The base_instructions.text field in Codex's session_meta contains the complete system prompt (15,225 bytes as of v0.117.0). This is standard Codex content, not user-specific. The previous 4KB read buffer captured only ~27% of the first line, producing invalid JSON that JSON.parse always rejected. Result: /retro global reported 0 Codex sessions for all users.

Test plan

• bun test test/global-discover.test.ts — 18 pass, 0 fail
• Manual verification: 174/174 historical Codex sessions now detected (was 0/174)
• Verified across all three originator types (codex_exec: ~15.9KB, Codex Desktop: ~16.2KB, codex_cli_rs: ~13.6KB)
• No private/user-specific data in the diff

🤖 Generated with Claude Code