feat(model): per-model context limits and pricing for the usage display

[justin-carper]

What

Populate Cursor models with per-model context window limits, per-token pricing, and output limits so opencode's TUI session header shows accurate tokens / context window % / cost.

Why

All Cursor models previously reported limit.context: 200_000 and cost: 0. The TUI session header rendered no context-window percentage and no dollar cost. Verified live (opencode 1.18.11): /config/providers returned "limit":{"context":0,"output":0} for all 33 models — the config channel (provider.<id>.models, built by toOpencodeModels()) emitted no limit/cost at all, and provider.models() hook output is discarded for providers absent from models.dev.

Changes

• src/model-limits.ts (new): shared per-model maps + longest-prefix resolvers for context limits (Cursor docs default windows, not Max Mode 1M), pricing (cursor.com/docs/models-and-pricing, USD per million tokens), and output limits. Imports nothing — avoids a circular import with model-discovery.ts.
• src/model-discovery.ts: OpencodeModelConfigEntry now carries limit and cost; toOpencodeModels() emits them on the config channel opencode actually reads. Cost uses the flat snake_case cache_read/cache_write keys the config schema requires (not the ModelV2 nested cache shape). Compile-time guards (_costKeyGuard/_limitKeyGuard) fail typecheck if the emitted keys drift from the accepted schema.
• src/plugin/model-v2.ts: buildModelV2Map() now imports the resolvers instead of duplicating them; stale doc comment fixed.
• Tests: 4 new config-channel tests (incl. a not.toHaveProperty("cache") shape trap guard), 7 new model-v2 tests. 405/405 passing, typecheck + build clean.

Verified end-to-end

Isolated opencode 1.18.11 server (published plugin excluded via XDG_CONFIG_HOME redirect) — /config/providers:

model	context	output	in/out	cache r/w
claude-opus-4-8	300000	64000	5/25	0.5/6.25
gpt-5.5	272000	64000	5/30	0.5/0
grok-4.5	256000	32000	0/0	0/0
claude-sonnet-4-6	200000	32000	3/15	0.3/3.75
claude-fable-5	300000	64000	10/50	1/12.5

Baseline before this branch: context: 0 for all 33.

Notes

• Cursor Models pool models (Grok 4.5, Composer 2.5, Auto) keep $0 — no per-token API charge.
• Values are a static snapshot; new Cursor model releases fall back to 200K/$0 until the maps are updated.
• Pre-existing (out of scope): a user provider.cursor.models.<id> override in config still wins over these defaults at plugin/index.ts:211 and would keep context: 0 for that model.