[chattymin]

Summary

Two documentation changes, no source changes.

1. Record the whole-file read class (docs/reference/defect-log.md).
#184 moved Codex rollout parsing to a chunked FileHandle stream. The class behind
that fix lived only in the PR description, which stops being reachable once the PR
scrolls out of view. The entry captures:

• why the whole-file read costs twice the file size, and why a per-line
  autoreleasepool cannot release the source string;
• the two mechanics that make the streaming version correct and bounded — splitting
  on the 0x0A byte (safe because UTF-8 continuation bytes are all >= 0x80) and a
  per-chunk autorelease boundary;
• a re-reading of perf: drain per-line autoreleasepool in usage-log parsing (peak -43%, steady -71%) #94: streaming was judged to regress memory because that chunk
  boundary was missing, so that judgement should not be cited to rule streaming out;
• the providers that still read whole files (parseClaudeFile, parseGrokFile,
  parseGeminiFile) and why converting them is deliberately deferred;
• that the regression guard is opt-in and therefore does not run in CI.

2. Add a reachability step to the defect protocol (CLAUDE.md).
Before adding or requesting a guard, confirm the scenario can occur in this codebase.
Injecting a defect proves a guard works; it does not prove the situation is reachable.
The worked example is the #181 review, where a test was requested for Codex root
folding and then withdrawn: there is no CODEX_HOME, so the roots are two hardcoded
siblings that cannot overlap, and the production callers are already normalized.

Type of change

• Bug fix
• New feature
• Refactor / cleanup
• Documentation

UI changes

None.

Checklist

• PR title and description are written in English
• No source changes — documentation only
• No copyrighted assets, secrets, or private tooling references are committed
• Every symbol cited in the new entry was verified to exist

Validation

• Cited symbols confirmed present: forEachCodexLine, parseClaudeFile,
  parseGrokFile, parseGeminiFile, CodexLargeRolloutPerformanceTests,
  POKETOKENBAR_RUN_LARGE_PERF.
• Measurements in the entry are from 59 real Codex rollouts (201 MB, largest 103 MB,
  release build): peak RSS 416 -> 66 MiB, parse 14.25 -> 2.16 s, byte-identical entry
  output before and after.