feat: badge model display names, configurable via options.badge

[BenjaminDSmithy]

What

Append a metadata badge to each model's display name so otherwise-identical builds are distinguishable in the OpenCode model picker, with the verbosity controlled by a new provider.lmstudio.options.badge option:

Tier	Result
off	Devstral Small
format	Devstral Small (MLX)
compact	Devstral Small (Q4_K_M · 14.3 GB · MLX)
full	Devstral Small (24B · Q4_K_M · 14.3 GB · MLX)

Default is full, so a user with no config sees exactly the same output as before this option existed — the change is purely additive. The option only dials the badge down.

{ "provider": { "lmstudio": { "options": { "badge": "compact" } } } }

Why

GET /api/v1/models already carries params_string, quantization.name, size_bytes, and format, but none of it reached ModelConfig. In practice an MLX quant and a GGUF quant of the same model share a display_name, and multiple quant levels collapse to the same row — the picker shows identical names with no way to tell them apart. This surfaces metadata the endpoint already returns; no lms CLI or SDK dependency, so it stays remote-server-safe. Raised against the rc.2 feedback tracker (#34).

How

src/plugin/enhance-config.ts:

• BADGE_TIERS / BadgeTier / DEFAULT_BADGE_TIER = "full", and resolveBadgeTier(raw) which fails open to full for any non-enum value (a typo must never break startup — matches the plugin's discovery-error-degrades ethos).
• modelBadge(model, tier) selects an order-preserving subset of [params, quant, size, format]. compact drops params on principle: params is identical across the quant/format siblings of one model, so it never disambiguates them — the tier keeps only the segments that do. (This avoids parsing display_name, which would be exactly the kind of name heuristic the README rules out.) Within a tier, each segment is still dropped when unreported.
• enhanceConfig resolves the tier once from existing?.options?.badge and threads it into toModelConfig. Separator is a single named constant (not a config knob — kept minimal). Only name changes; the model key/ID is untouched.

src/types/index.ts: adds params_string and size_bytes to LMStudioModelSchema, both .nullable().optional() so older servers and the existing fixture still validate.

The badge key and forwarded options

badge is a plugin-only option, and it is deliberately not stripped from the options forwarded to @ai-sdk/openai-compatible. Two reasons, both verified against the bundled versions:

1. Safe: createOpenAICompatible reads only named fields (baseURL, apiKey, headers, queryParams, fetch, …) and ignores unknown keys — badge is never placed in a header/URL/body, never sent on the wire, never throws.
2. Idempotent: options is the one config record that persists across repeated config-hook invocations (it's how apiKey survives re-runs). Stripping badge would revert the user's tier to the default on the second config load. An enhanceConfig-twice test asserts both the option's persistence and that the chosen tier is still applied on the second run.

If you'd rather strip it anyway, I'm happy to add a persistent side-channel for the tier — but not-stripping is the simpler correct choice here.

Contract note

The rc.2 contract (#34) says "display names come directly from LM Studio metadata." This decorates the display name; the key is still verbatim, and every segment is LM Studio metadata. docs/v1-contract.md is updated with the badge mapping row and the default. If you'd prefer the default to be off (badge fully opt-in) rather than full, that's a one-line change — say the word.

Tests

• modelBadge per-tier (incl. intra-tier segment drop) + formatSize boundaries + resolveBadgeTier fallback.
• enhanceConfig integration: off yields bare names, compact applies + survives a second run (no-strip idempotency), invalid value falls back to full.
• 29→32 pass; tsc --noEmit clean; eslint clean; lockfile untouched.
• Independent second-opinion review (codex, different model family): no P0/P1/P2 issues; confirmed no empty-parens/dangling-separator cases, the no-config regression is nil, and the no-strip decision is safe + idempotent.

README (Model mapping), CHANGELOG.md (Unreleased → Added), and docs/v1-contract.md updated.