fix(models): sync bundled offline fallback data with models.dev (2026-09-08)

[ltmoerdani]

Summary

Full audit of the bundled offline fallback tables against the live models.dev registry (api.json, fetched 2026-09-08) found three drift classes. All of these tables only matter when the live models.dev fetch and the gateway /models endpoint are unavailable, but some wrong values were large enough to matter anyway (up to 6x cost under-reporting).

Every drifted value was traced with git log -S before changing it, to separate stale snapshots from deliberate deviations.

1. Context limits (src/models/modelTables.ts)

Model	Before	After	Why
kimi-k2.7-code	256,000	262,144	Verified against models.dev on 2026-06-15 (issue #25); models.dev changed since
minimax-m3	512,000	1,000,000	2026-06-09 snapshot; Zen still lists 512K, Go was raised to 1M
qwen3.6-plus	262,144	1,000,000	Same pattern; Zen still lists 262K
qwen3.7-plus	262,144	1,000,000	Copy-paste from qwen3.6-plus in 1c26fd1 (2026-08-13), never verified

The key signal: models.dev now differentiates Go vs Zen per model, and the bundled Go values exactly matched the Zen ones. That is the same cross-provider contamination pattern fixed in docs/issues/02.

Kept as-is (deliberate, verified against the gateway): deepseek-v4-flash output cap 131,072 (#171) and Go minimax-m2.5 output 65,536.

2. Usage tracker pricing (src/usage/pricing.ts)

The static GO_MODEL_PRICING fallback had drifted badly from the registry:

Model	Bundled	Actual	Impact
mimo-v2.5-pro	$1.74 / $3.48	$0.435 / $0.87	4x over-report
deepseek-v4-pro	$1.74 / $3.48	$0.66 / $1.98	~2.6x over-report
mimo-v2-omni	$0.14 / $0.28	$0.40 / $2.00	big under-report
deepseek-v4-flash	$0.14 / $0.28	$0.22 / $0.66	under-report
minimax-m3	$0.60 / $2.40	$0.30 / $1.20	2x over-report
kimi-k3	missing (fallback $0.5/$2.0)	$3.00 / $15.00	6x under-report

Table rewritten from the current Go snapshot: 15 new models added, minimax-m2.1 / minimax-m2 removed (no longer served), hy3-preview renamed to hy3 with the corrected price.

3. Go fallback catalog (src/provider/definitions.ts)

fallbackModels still listed the retired hy3-preview, omitted kimi-k2.7-code and everything added after June, and included deprecated models, contradicting the README's deprecated-filtering promise. Refreshed to the 27-model curated active set.

4. Stale tests

metadata.test.ts, registry.test.ts, thinking.test.ts still referenced hy3-preview (passing only because the routing catch-all treats both ids the same). Updated to hy3. The goUsageTracker and metadata unit-test locks that pinned the old pricing/context values were updated as part of the sync.

Verification

• npm run lint: all 7 checks pass, 462/462 unit tests
• models.dev snapshot diffed programmatically before/after (curl + python3)

Caveats

• The 1M context limits were not verified against the live gateway with oversized requests. models.dev is OpenCode's own registry, but [BUG] deepseek-v4-flash sends max_tokens=384000 but OpenCode Go now caps completion at 131072 #171 showed the gateway can lag the registry, so flagging it here.
• DEFAULT_INLINE_MODEL is still qwen3.5-plus, which is upstream-deprecated. Left unchanged; worth a follow-up.

Doc: docs/issues/100-20260908-bundled-model-data-sync.md