c639d60859
`POST /outcome` is the only ground truth this router has, and feedback.py is the one loop with no timer -- poller, seed sweep, backup and offsite all have one. Live state on 2026-09-15: `proficiency` last written 2026-09-10, with 404 unapplied attributable outcomes and thousands of decisions routed off the stale scores in between. The fold is irreversible. add_outcome accumulates into a running mean and recompute_category re-derives every row in the category from the new peer rate; neither keeps the pre-fold value anywhere, and verifications.applied_at means a second run will not redo the work either. So the deliverable is a preview plus units that ship unstarted, not an automatic fold. `--dry-run` now projects instead of describing. feedback_preview copies the database into memory, runs the REAL add_outcome against the copy, and diffs the two proficiency snapshots. It does not re-implement the empirical-Bayes conversion -- a second implementation would drift, and a confidently wrong forecast of an irreversible action is the worst failure available here. Two kinds of movement come out, and the second is the surprise: `direct` rows carry new outcomes of their own; `ripple` rows carry none and move anyway, because the whole category is re-derived against a peer rate the new evidence just changed. On the live backlog 404 samples across 17 pairs move 17 rows directly and 363 by ripple, so ripple is digested per category and `--csv` carries every row. The units are named and hardened like the poller/seed pair, and take no EnvironmentFile and no network-online.target because the fold makes no HTTP request of any kind. The .service has no [Install] section, so it cannot be enabled on its own -- "not enabled by default" is structural rather than a README promise. 12h cadence because the fold is exactly additive: ten folds of five land on the same numbers as one fold of fifty, so cadence caps staleness and batch size and cannot change where the scores end up. Three corrections to what was in the tree: - The timer carried `Persistent=true`, which systemd.timer(5) says "only has an effect on timers configured with OnCalendar=". This timer is monotonic, so the line bought nothing; OnBootSec is the real catch-up. The sibling poller and seed timers carry the same inert line -- noted, not fixed in passing. - `--csv` without `--dry-run` was accepted and discarded, and the run it was silently dropped from is the irreversible one. It is now refused. - preview() copied the whole 41 MB database into memory just to print coverage, then project() copied it again. Coverage only reads, so it takes a mode=ro handle instead. Tests: 2068 -> 2087. The load-bearing one asserts a dry run leaves the database byte-identical -- proficiency rows, verifications.applied_at, and the file's sha256 -- while still reporting the 8-sample delta it would apply. Verified non-vacuous by handing copy_database the real connection and watching it fail. A second test folds for real onto an identical copy and demands the projection match every score, which is what stops the preview drifting from the store. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U
1.5 KiB
1.5 KiB