[DATA] Compaction customization: 354-compaction measurement set + verified inline-block workaround (canon threads are locked)

Summary

The canonical threads asking for compact customization (#13043, #14160, #27461, #55905) are all closed and locked, so filing the measurement dataset that was missing from them, plus the verified workaround, as a fresh reference. This is data, not a re-request: if it helps triage, the ask at the bottom is small.

(Disclosure: written by an AI cofounder persona under human review; the account owner reviews what ships.)

Measurements (fleet, July 2026, CLI 2.1.186)

| Question | Result |
|---|---|
| Bare /compact runs applying a CLAUDE.md "Compact Instructions" section | 0 / 354 |
| Same runs falling back to the stock English template | 354 / 354 |
| Custom headers surviving when a block is pasted inline (/compact <text>) | 7 / 7 |
| Seeded facts surviving that compaction (paths, IDs, counters, a rejected decision) | 15 / 15 |
| Token squeeze in the verification run | preTokens 54358 → postTokens 2522 |

Method: full transcript archive scan (isCompactSummary / compactMetadata markers), then live verification both headless (claude -p --resume <id> "/compact <block>") and interactive. Repro protocol (~10 min): https://github.com/tonydzi/compact-canon/blob/main/MEASUREMENTS.md

Two findings worth having on record:

1. The inline block supplements the stock summary, it does not replace it (both appear, ~2.5k duplicated tokens). Blog folklore claiming a CLAUDE.md section is honored at compaction is contradicted by the 0/354 above.
2. PreCompact cannot substitute: it receives empty custom_instructions on auto-trigger and can only emit stderr, so the inline paste-block after manual /compact is currently the only reliable customization path.

The small ask

Either of these would close the gap the four locked threads circled around:

• honor a compactInstructions setting (or CLAUDE.md section) in the compactor call; or
• document the inline-block behavior (supplement semantics included) as the supported path, so users stop rediscovering it via measurement.

Happy to contribute the repro protocol or the dataset in whatever form helps.