feat: opencodeModel: "inherit" — capture with the session's own model

[anagnorisis2peripeteia]

Motivation

Pinned opencodeModel ids rot: when a provider renames a model (or a local endpoint changes its served model id), auto-capture silently fails at session.create on every prompt until someone notices the log spam. That is how I found this — capture had been broken for days on a stale local-model id. Separately, users who switch models during the day may prefer each conversation to be captured by the model that actually produced it.

What this does

"opencodeModel": "inherit" makes auto-capture use the exact provider/model opencode resolved for the captured prompt:

• a chat.params hook records input.model.providerID / input.model.id onto the prompt's existing user_prompts row (two new nullable columns, added with the same duplicate-column-guarded ALTER TABLE pattern as capture_attempts), keyed by message.id — the same key capture already uses, so the mapping is exact rather than inferred from transcript position;
• at capture time, "inherit" resolves provider/model from the stored row; the existing isProviderConnected guard runs against the resolved provider;
• if a prompt has no recorded model (predates the upgrade, or the provider path didn't invoke chat.params), capture throws and the existing retry/cap machinery handles it — it never guesses.

Pinned configuration behaves exactly as before; the hook returns immediately unless "inherit" is configured.

Notes for review

• opencodeProvider must still be set (the existing opencodeProvider && opencodeModel gate is untouched for diff minimality); with "inherit" its value is only used as a gate. Happy to change the gate if you prefer.
• On opencode v1.17.11 I observed chat.params firing reliably for models served through the standard providers, but not for every custom npm-provider path — the hard-error fallback covers those prompts. Worth knowing for docs/support.

Testing

• tsc --noEmit and Prettier clean; migration verified against an existing populated user-prompts.db (old rows read back with null model fields).
• Live-tested against opencode v1.17.11: sessions on a local llama.cpp-served model recorded provider_id/model_id correctly on their prompt rows, and capture consumed the recorded model (verified via capture-session behavior/logs).
• Rows without recorded models fail capture with the new explicit error and are retried/capped by the existing machinery.