Reduce idle Codex adapter CPU usage

[edwin-zvs]

Summary

• tail Codex root and native-child rollout JSONL files from byte offsets instead of rereading full history every 500 ms
• cache rollout metadata after reading only the first JSONL record
• discover new rollouts from changed directory mtimes, with a periodic full-scan safety net
• preserve full native-child backfill, partial-write retries, truncation recovery, and /clear rebinding

Motivation

A live sample of an idle construct __adapter codex process showed CPU in recursive rollout discovery and whole-file read_to_string calls. On this machine, 28 Codex adapters consumed roughly 33–36% of one core in aggregate. The rollout store contains 2,120 files / 1.0 GB; one sampled adapter reread an 18.9 MB root transcript multiple times per second while idle.

After this change, idle root and child transcripts cost a file-length metadata check with no content read. Directory discovery normally stats the 68 rollout directories and only enumerates a directory whose entries changed; the 10-second full enumeration backstop handles coarse or coalesced timestamps.

Verification

• cargo test -p construct-adapter-codex — 31 passed
• cargo clippy -p construct-adapter-codex --all-targets --no-deps -- -D warnings -A clippy::too_many_arguments
• cargo build

No screenshot or recording: this is an internal adapter performance change with no UI difference.