Add scratch-file offload for elided tool output (ports OpenDev's truncation.rs pattern)

[maxkilla]

What

Ports a pattern from OpenDev's truncation.rs (opendev-to/opendev, crates/opendev-tools-impl/src/truncation.rs) into comb: before eliding the middle of oversized tool output, save the full original to disk so it's recoverable.

Why

comb's compress() currently elides the middle of oversized output permanently — there was no way to get it back if it turned out to matter. OpenDev solves the same problem in its own tool-output truncation path by always writing the full output to a scratch file first and returning a path alongside the truncated preview.

What changed

• saveOverflow(text): writes the full untouched text to ~/.claude/comb/tool-output/tool_<timestamp>_<8-hex>, mode 0o600. Never throws — on any FS error it returns null and compress() just omits the recovery hint; the file write is best-effort, the compression itself is not.
• 1MB cap on the saved file (MAX_OVERFLOW_BYTES), matching OpenDev's constant. Past the cap, the saved file is itself capped head 75% / tail 25%, not a flat cutoff — so both how the output started and how it ended survive even in the worst case.
• compress()'s marker now includes , full output: <path> when a save succeeded.
• COMB_COMPRESS_SAVE_FULL=0 opt-out. Defaults on (opt-out rather than opt-in) since silently losing tool output seems worse than an occasional stray file on disk.
• cleanupOldFiles(): 7-day retention sweep. Exported but not auto-invoked by the hook itself — PostToolUse runs in the hot path on every matched tool call, so a directory scan there is wasted work every time. Meant to be wired into a SessionStart hook or cron if disk growth in OVERFLOW_DIR becomes a problem.

Testing

• All 19 pre-existing tests in test/compress.test.js pass unmodified — nothing about the existing elision/salvage/gate/rule-store behavior changed.
• 9 new tests in test/overflow.test.js:
  • saveOverflow writes the exact full text and returns a path
  • the 1MB cap actually caps, with head/tail both surviving
  • compress()'s marker contains a working path, and the saved file matches the original byte-for-byte
  • COMB_COMPRESS_SAVE_FULL=0 writes nothing and omits the path from the marker
  • cleanupOldFiles() keeps recent files, removes files past retention (verified via fs.utimesSync, not a real 7-day wait), leaves non-tool_* files alone, and returns 0 gracefully when the directory doesn't exist yet
• Also ran the hook end-to-end via actual stdin/stdout (as Claude Code would invoke it) with a synthetic oversized Bash tool_response, confirmed the returned JSON's updatedToolOutput.output contains a working path and the file on disk matches the pre-elision original exactly.

What I deliberately did NOT port from OpenDev

• No line-count threshold alongside the existing character-count threshold — comb's HEAD_CHARS/TAIL_CHARS/THRESHOLD are already purely character-based, and OpenDev's dual line+byte trigger wasn't something comb's actual design was missing; would've been scope creep.
• Kept COMB_COMPRESS_SAVE_FULL opt-out (vs. no toggle at all in OpenDev) since comb already has a convention of explicit env-var toggles for anything that touches the filesystem/network (see COMB_RULE_STORE_URL).