Registry maintenance: weekly models.dev diff report and monthly reachability probe

[cunninghamcard-bit]

Part of #166 (track A follow-up; RFC-0033 §4.1 items 4–5).

provider_registry.json is the hand-maintained source of truth (251 rows) and must stay that way: it carries manual fixes (94b5000, b24e59e) that any regeneration would clobber. What is missing is a way to notice when upstream reality moves. A one-off comparison against models.dev on 2026-09-04: 97 rows exist on both sides, 27 of those disagree on base_url, and 116 models.dev providers are absent from aimux.

Reference: pi-ai regenerates model data from models.dev at build time and fails the build when the data is stale (scripts/generate-models.ts, scripts/check-model-data.ts). aimux cannot regenerate, so the equivalent is a diff report plus a reachability probe.

Deliverables

1. scripts/sync_registry.py --report — fetch https://models.dev/api.json (213 providers, 187 with base URL, all with env names); optionally read a local reference/litellm checkout (absorbs extract_litellm_bases.py / scan_litellm_urls.py, which are then deleted). Output three tables: providers missing from aimux, base_url disagreements, env_var disagreements. Rows with a "note" field are treated as intentional deviations and not reported.
2. scripts/probe_registry.py — for each row, unauthenticated GET {base_url}/models; 401/403 counts as alive; DNS failure or 5xx is reported. Never edits the JSON.
3. A scheduled GitHub Action: weekly sync report, monthly probe, posting to a fixed tracking issue (or opening a PR with the proposed row edits). Non-blocking.
4. Add the optional note and status ("unreachable") fields to the row schema and to the Rust validation tests in provider.rs.

Acceptance

• Running the report locally reproduces the 27 / 116 numbers above (until the triage issue lands).
• CI on pull requests is unchanged; the schedule never blocks merges.