Pruned from STATE.md, 2026-09-19

STATE.md is capped at ~200 lines because it is the first thing a cold session reads and a long one stops being read. These blocks were true and are now history; the LESSON from each was kept in STATE.md in one line, the narrative is here.

samo-dev drifted while STATE.md said it was in step (2026-09-13)

samo-dev had drifted four migrations behind while a line in STATE.md asserted it was current. Nobody was lying: the line was written when it was true and never re-checked, which is what a decaying fact does in a document with no compiler.

Two things changed as a result, and both are mechanisms rather than sentences: a merged migration nobody applies now shows up in the pull request itself, and npm run deploy:owed asks PRODUCTION before it gives a verdict instead of reading a line.

The rule that survives: ask npm run migrate:status -- --dev, never the prose. Note the -- — a bare --dev is swallowed by npm and the command then reports PRODUCTION, which is its own entry in the mistakes files.

Why it happened, in full: docs/state/phuriphatma.md. What is still owed on the dev project: docs/state/HANDOFF.md §8.

The roster import's own numbers (2026-09-14)

The import landed 1,611 students and 165 held rows in one transaction, with 306 สาย and ten houses. Those numbers have all moved since — held is smaller every time a ฝ่าย answers, students larger — which is exactly why STATE.md no longer carries them and npm run handoff:check fails when a document contradicts the database.

advisors being empty is NOT historical and stayed in STATE.md: the two rows it once held were test data pointing at สาย 100/200, which are real สาย in the file, so twelve students briefly saw an อาจารย์ who does not exist. Snapshot and write-up: docs/mistakes/postgres-schema.md.

The migration paragraph STATE.md carried until 2026-09-19

One 1,900-character line listing what every migration from 0183 to 0194 did. It was accurate and it was in the wrong place: STATE.md is read first by every cold session and is capped at ~200 lines for that reason, while this is reference material nobody needs in the first minute. The proofs are the authority for all of it; npm run migrate:status is the authority for what is applied.

• Migrations through 0194. ALL 43 LIVE PROOFS GREEN (2026-09-14). 0194: an import may FILL the registry, not overwrite it — the old guard skipped the write entirely, so 136 people had a NULL ชื่อ in people while ระบบบ้าน had it; 20/20, watched failing first. Registry now 0-diff against BOTH placements (§F of that proof asks it live). 0193 keeps what the handover file said in a cell the cleaner emptied — a removed address and the รุ่น of a row with no รหัส — as EVIDENCE the admin can see and nothing can resolve a row with; 18/18, watched failing first. 0183/0184 are the PORTAL half of Discord role sync — discord_links, team_nodes.discord_role_id + discord_role, plus discord_role_targets(), the ONE function that decides what a person is due — applied and proved 23/23 + 18/18, both watched failing first. The seed ticked 107 of 299 nodes; the rest are the owner's review. 0187 closed a hole where UNLINKING kept your ฝ่าย roles for ever — §14b, its one home. HANDOFF §11 and §13c both CLOSED. 0188–0190 keep the import lines that have no kkumail and let the student claim their own seat with รหัสนักศึกษา + ชื่อ — 33/33, each fix watched failing first. 0188's table was born anon-writable (a pg_default_acl on public grants anon arwdDxtm to every NEW table; every other ระบบบ้าน table revokes it per object). 0189/0190 came out of reviewing 0188: a claim reached students without last_import_batch, so both mirror triggers took the NON-import branch and the file's spelling overwrote a curated registry name silently.