fix(velocity): extend canonical model format to support experimental labels like "owl-alpha"

[behnamnba1900-art]

Resolves #1193

Decision taken: chose path 2 (intentional non-Claude / experimental model support) rather than editing the source DB row.

Changes:

• Extended CANONICAL_MODEL regex in scripts/velocity-log.js and scripts/error-log.js (and the report test) to /^[a-z]+-[a-z0-9.]+$/.
  This accepts both the original <family>-<major>.<minor> (e.g. opus-4.8, sonnet-4.6) and labels like owl-alpha.
• Updated error messages, inline comments, and the model column docs in docs/puzzle-velocity.md + docs/errors-schema.md.
• The historical "owl-alpha" row (CSV row ~1010, from incaberry session) is now considered valid under the convention. No data deletion or DB edit required.
• The velocity-csv test now reports 0 non-canonical models for this case (previously would have warned).

This makes the standing red / drift go away while keeping the notice-not-prevent philosophy from #1184 for future model drift.

Tested locally: the specific test logic now accepts the row; no other non-canonical models were present.

[behnamnba1900-art]

Checked out #1193. The "owl-alpha" entry was the one triggering the model naming violation in the puzzle-velocity-csv test (even after the notice-not-prevent change from #1184).

I went with extending the canonical format instead of trying to scrub the historical row from the DB/CSV. Updated the regex in velocity-log.js and error-log.js (and the report test) to also accept - forms like "owl-alpha" alongside the versioned ones. Synced the docs and now the current data passes cleanly with zero non-canonical models reported.

Felt like the right long-term call given the existing comments about models being an open-growth list. If you'd prefer to treat it as bad data and clean the row instead, just say the word and I can adjust.

Closes #1193