feat(llm): add Cohere chat adapter (command-r / command-r-plus)

[mvanhorn]

What

Adds a Cohere Chat API adapter so users can route the watcher to command-r / command-r-plus instead of being limited to Anthropic + the OpenAI-compatible surface.

Cohere's wire format differs from OpenAI's:

• Request: { message, preamble, chat_history? }
• Response: { text, meta: { tokens: { input_tokens, output_tokens } } }
• Error: { message: \"...\" } (flat, no nested error object)

The adapter maps systemPrompt -> preamble and transcriptPrompt -> message, matching how the watcher uses every other adapter (single system + single user turn per call). chat_history is omitted because the watcher always sends a fresh transcript excerpt; multi-turn history is the caller's responsibility.

Files

• cohere.ts (new): adapter using /v1/chat, Bearer auth, classifyHttpError, and a Cohere-specific error message extractor
• types.ts: \"cohere\" added to ProviderKind
• registry.ts: getAdapter routes kind === \"cohere\" to createCohereAdapter; defaultBaseUrl returns https://api.cohere.com
• pricing.ts: adds command-r (15/60c per 1M) and command-r-plus (250/1000c per 1M)
• list-models.ts: adds a Cohere branch hitting /v1/models with Bearer auth (the existing .models fallback in the parser already handles Cohere's response shape)
• llm.test.ts: 5 new adapter tests + 2 pricing tests

Test plan

• bun test src/server/services/ai/llm/llm.test.ts: 20 pass / 0 fail (52 expect() calls, was 13 / 0 before)
• bun run typecheck: 0 errors (the new ProviderKind exhausted a previously-incomplete switch in list-models.ts; that's now fixed)
• bunx biome check src/server/services/ai/llm/: 0 errors

Notes

• completeStream is intentionally not implemented in this PR. Cohere streams via SSE-style event: text-generation chunks; the existing streamWithFallback wrapper in types.ts will serve full responses as a single delta until that's added in a follow-up.
• extractCohereErrorMessage parses the { message } shape and falls back to the raw body on parse failure, so opaque error responses still produce useful LlmError.message text.

Closes #7

[jstuart0]

Thanks for this — really appreciate the work, and apologies for the timing churn. After you opened this, we landed a refactor on main (commit 781f2b0, slice TYPE-2a) that consolidates ProviderKind into a single KNOWN_PROVIDER_KINDS as const declaration in src/shared/types.ts. The runtime allowlist in routes/ai.ts, the web-side AiProviderKind type, and the UI dropdown's exhaustiveness guard all derive from it now — so adding a kind only requires updating one place rather than four.

Net effect: your diff gets smaller after a rebase.

Concrete checklist for main:

1. src/shared/types.ts — add "cohere" to KNOWN_PROVIDER_KINDS
2. src/server/services/ai/llm/cohere.ts — your adapter (unchanged)
3. src/server/services/ai/llm/registry.ts — your getAdapter case + defaultBaseUrl entry (the existing exhaustive switch will require this to compile)
4. src/server/services/ai/llm/list-models.ts — your Cohere /v1/models branch (existing exhaustive switch likewise)
5. src/server/services/ai/llm/pricing.ts — your two MODEL_RATES rows (kept the command-r-plus before command-r ordering — nice catch on first-match)
6. src/web/components/settings/AiSettingsPanel.tsx — add an entry to PROVIDER_KIND_OPTIONS. The as const satisfies exhaustiveness check will fail compile if you forget this step.
7. src/server/services/ai/llm/llm.test.ts — your tests (unchanged)

You can drop the changes to src/server/routes/ai.ts:494-500 (the local PROVIDER_KINDS array no longer exists — the runtime check is now KNOWN_PROVIDER_KINDS.includes(...)) and to src/web/lib/api.ts:1123 (AiProviderKind is now a re-export of the shared type).

Once rebased, all your existing 7 tests stay green, plus tsc will flag any wiring point you miss — a strict improvement over the state this PR was originally written against.

Happy to review the rebase when you push it, or take it from here ourselves if you'd rather not deal with the churn — just say the word. Either way, thanks again.