fix(models): retire dead model ids, reconcile picker with gateway, fix franklin models pricing

[VickyXAI]

Reported as "Franklin models still list old models." It was structural, and it turned out to be two bugs.

1. The picker never talked to the gateway

The in-session /model picker rendered a hand-maintained array, so it drifted on every upstream change. (franklin models was live-fetching — two surfaces, one stale.) gateway-models.ts already existed with a live catalog client whose docblock claims it "replaces the hardcoded pricing/model tables" — it didn't; only the media tools imported it.

It now reconciles via reconcilePicker(): rows the gateway stopped listing drop themselves, prices refresh live, curated labels/order stay (gateway names overflow the tuned columns), and it falls back to the static list when offline.

Dead ids

Verified by probing the chat endpoint — 402 = exists, 400 = unknown. The /api/v1/models listing is not authoritative: it hides working models (grok-3, gpt-5-nano, opus-4.6, kimi-k2.6 are all absent yet alive). A naive diff against it produced five false "dead model" findings that were caught and discarded.

id	impact
anthropic/claude-haiku-4.5-20251001 → 400	9 sites, incl. compact.ts — selected for every Sonnet/GPT-5.4/5.5/Gemini-Pro session, so long sessions compacted against a model that 400s
openai/gpt-5 → 400	tool-use fallback chain

Both stay in pricing.ts/tokens.ts — those are keyed by whatever id a past session recorded, and the file already retains retired ids for that.

Maverick doesn't 410 — it lies

nvidia/llama-4-maverick left the catalog but still answers 200, silently served by another model (probed live: it and the free default qwen3-next both came back as nemotron-3-super-120b-a12b-free). That made the free "diverse-family" fallback chains fake — primary and fallback resolved to the same backend, so the fallback could never rescue a turn the primary failed. Replaced with mistral-nemotron, verified to serve itself. Also removed from the vision allowlist, which contradicted routeRequest()'s own "maverick is text-only" comment — free vision turns were routed to a text-only model.

Adds Grok 4.5 (the gap that prompted the report): the gateway calls it xAI's flagship, so grok follows it per the bare-alias-tracks-flagship convention. 4.3 stays pinned — cheaper, and 1M ctx vs 4.5's 500K.

2. franklin models priced the whole media catalog at $0.00

It modelled every model as { input, output }, so the entire image/video/music/speech catalog (22 paid models) printed as $0.00/M under a "Paid Models" heading. On a price list, that's the worst direction to be wrong in. The gateway bills seven different ways; each now renders in its own unit, and unknown modes are still listed with raw pricing rather than a wrong one.

Same root cause, latent: per_character and per_generation are real billing modes but were missing from BillingMode, so estimateCostUsd fell through the switch and returned 0 — a paid ElevenLabs call estimated as free. No speech tool calls it today, so nothing leaked; fixed before one does.

Verification

• 562/562 tests pass (8 new: 5 reconciliation, 3 pricing/format)
• Picker driven end-to-end; offline fallback exercised with a stubbed fetch
• All 20 models the picker now offers probed live — every one resolves
• franklin models output verified live against all 78 models
• Both commits are green independently (bisect-safe)