Correct two stale claims in tiers_ratchet.py's docstring

[andrewboudreau]

Docstring only. --check still PASS 2565/2565; no behaviour change.

Both claims were introduced by me, and both actively mislead a reader:

1. "NOT WIRED INTO ANYTHING YET -- deliberately."

False since #2011, which added .github/workflows/converted-ratchet.yml (pull_request + push: main). An agent working #1993 read this line, concluded the gate was unwired, and spent a cycle on a theory that could not be true. Replaced with what the workflow actually does — and it points at that workflow's on: block as the authoritative path list rather than duplicating it here, since a duplicated list is the thing that goes stale next.

2. "staleness only ever runs one way."

False for two shapes, both of which cost real time this week.

• TU promotion removes banked paths. classify_missing() already handles it — resolves the gone path through tools/tu_manifest.py and reports MOVED; when the absorbing file is itself CONVERTED the removal is absorbed_clean, needing no --reason and logging no exception row. The banked COUNT drops, and that is correct. Stated explicitly because I circulated the opposite rule — "the floor is 2565, never go below it" — to three agents, and two of them had to come back with the source to correct me. This is a SET ratchet. There is no floor number to defend, and a promotion batch is expected to lower the count.

• A plain file rename has no detection at all. It reports GONE -- not a tracked source file any more and demands a --reason, which would write a permanent fake backslide row into config/converted-backslide-exceptions.jsonl. The workaround (apply the renames to the parent's banked set first, then regenerate) is now recorded in the docstring until real rename detection lands.

Also records the failure shape from #1996: a CI red here is usually a base desync, not a regression — a PR cut before a re-bank merges against a baseline that predates it, inherits the red, and the next unrelated src/** PR is the one that turns red for it. Fix is to merge main in, never to lower the pin.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WdCK1xgrJdiJzPCh3bAJfQ

[tangos-validator]

✅ PR validation — Passed

noverify: no source/build-data changes in this PR

_{Each changed src/*.c|*.cpp is compiled and its relocated bytes compared to the binary data on a private build box. Passing requires every changed file to reproduce the ROM byte-for-byte with correct relocation targets — this catches WRONG-DEST relocations and non-reproducing near-misses that ledger-scoped linkcheck skips.}