feat(providers): re-query effort options after model switch

[j35dev]

OpenCode advertises its effort selector per-model over ACP (reasoning-capable models only), so the default-model catalog probe hides the chip even with a reasoning model selected. This adds a model-aware lookup plus turn-time refresh, all through ACP. pnpm verify green.

---

[j35dev]

Review — request changes

The ACP adapter half is sound: applyModel returns post-switch configOptions, and the new driver test covers a thinker model advertising low/medium/high and applying saved high.

The renderer EffortChip does not finish the job.

1. Unsupported models keep a stale chip (SessionView.tsx ~1083)
Empty providers.efforts is ignored, so a plain model keeps the previous reasoning model's options. Empty is overloaded (no selector vs probe failed). A successful empty lookup should setOptions([]) and hide the chip. setDynamicEfforts from a model-specific probe also poisons the kind-global catalog.

2. Startup / model switch can wipe a valid saved effort (SessionView.tsx ~1061)
apply() from providers.models sets loaded=true with empty kind-catalog options before the model-specific RPC returns. The validate effect then session.updates effort: null. For OpenCode with persisted effort=high, that value is gone by the time the chip appears.

loaded is never reset on modelId change, so the same hole fires on every switch.

The one new renderer test is happy-path only (empty catalog, non-empty model probe, chip shows). It never passes a saved effort.

Do not merge until 1 and 2 are fixed and tested. #224 (composer picker collapse) is a subset of this branch and is landing separately.

Skip list: #223 stays open as requested.

[j35dev]

Update: #224 landed on main (rebased through #222). This branch still contains that commit plus the effort-requery work, so it will need a rebase. The two EffortChip bugs in the review above still block merge.