The spec's own example IDs would have failed the checksum

Phase 2 Session 3 closes out with 161 tests passing and a small, slightly embarrassing discovery: every example participant ID in the spec, the README, and CLAUDE.md was illustrative in the worst sense. They looked like valid LooMed IDs. None of them would have passed the validator I wrote this session, because the checksum in every one of them was made up.

A checksum is not a format check

ParticipantId::new() previously did the minimum: check the type prefix, check there were enough --delimited segments, done. Spec §3.1 asks for more — every ID ends in a two-character checksum, a CRC-8 (polynomial 0x07, no reflection) computed over every segment that precedes it and rendered as two Crockford Base32 digits. That’s not a cosmetic addition. A format check catches a garbled string. A checksum catches a correct-looking string with one wrong character — the case that actually matters, because that’s what a mistyped or partially-copied participant ID looks like in the wild.

The test that made this concrete: take a valid clinician ID, LMD-APL-3NKWQ7HZRC-5N, and mistype the scope segment — APL to APM. Same length, same alphabet, still looks plausible. The checksum, computed over type + scope + base32 ID, catches it immediately. That’s the point of hashing the preceding segments instead of just the random ID portion: a transcription error in the institution’s scope code is just as dangerous as one in the random segment, and the old validator would have silently accepted it.

Regenerating the fixtures

Once compute_checksum existed, I ran it against the five illustrative IDs already sitting in the spec’s own “Examples” section — LMP-7XKQR2MNVB-F4, LMD-APL-3NKWQ7HZRC-8A, and so on. All five failed. They’d been typed by hand at some point in an earlier phase, formatted correctly, never actually run through a checksum algorithm because the algorithm didn’t exist yet. Every fixture across loomed-core, loomed-cli, loomed-store, the README, and CLAUDE.md needed the same treatment: regenerate against the real function, not eyeball a Base32-looking string.

Two smaller, quieter wins

loomed log now prints a one-line typed summary under each commit — FBG: 98.5 mg/dL, Metformin 500mg × 30d — built from the six clinical payload structs added in the last session. It fails soft: an empty payload or a mismatched type just omits the line rather than dumping raw JSON. And loomed show now surfaces pre_sync_previous_hash and pre_sync_commit_id when a commit has been through Sync Rebase, so loomed show finally tells the whole sync story instead of the parts that predate the fork.

Crate	Tests
loomed-core	60
loomed-crypto	17
loomed-store	18
loomed-sync	15
loomed-cli	51
Total	161, 0 failures

Next is Phase 3: consent tokens, and the first release plan that separates “what ships in v1.0” from the full phase roadmap.