fix(models): per-provider discovery state, Fable 5.1 in the picker, Codex effort from the CLI

[bbsngg]

Follow-up to #213 / #216 for three problems found while using the picker on main.

1. Provider switch corrupted the saved model (bug)

useHarnessModels set its state asynchronously, so for one render after a provider switch it still held the previous provider's list while provider already named the new one. ChatComposer's rescue effect then saw the Claude model "missing" from Codex's list and replaced it with Codex's default: localStorage['claude-model'] ended up as gpt-5.6-sol, and every new Claude session would have failed.

• The hook now tags each answer with the provider it was fetched for and only exposes an answer for the current provider.
• useChatProviderState drops a stored model that cannot belong to its slot (a gpt-* under claude-model, a claude-*/alias under codex-model) and falls back to the default, so anyone already affected recovers on next load. Verified in the browser: the corrupted value was replaced by claude-fable-5.

2. Fable 5.1 was not selectable

The Claude probe runs with settingSources: [] so no hooks fire. That also hides the model configured in ~/.claude/settings.json, and that is exactly where the CLI's own /model menu gets Fable 5.1 from (0.3.226's base menu lists Fable 5, not 5.1). Discovery now reads model from the user settings file (honouring CLAUDE_CONFIG_DIR) and ANTHROPIC_MODEL, and offers it the way the CLI does. claude-fable-5-1 / claude-fable-5-1[1m] are also in the built-in list so they are available even with discovery off. Duplicate CLI display names ("Fable" for both 5 and 5.1) are disambiguated from the description.

Real probe output on this machine after the change: default, opus[1m], claude-fable-5[1m] → Fable, sonnet, haiku, claude-fable-5-1[1m] → Fable 5.1 [1M], then the built-ins.

3. Codex reasoning-effort selector missing for discovered models

Support was a table keyed by exact model name; gpt-5.6-sol, gpt-5.4-mini, gpt-5.3-codex-spark fell to "default only" and the selector disappeared. Codex's model/list reports supportedReasoningEfforts and defaultReasoningEffort per model, so discovery now carries them through and the selector, its gating, and the saved-effort normalisation use that list first, with the table as fallback. The ultra effort the CLI offers (gpt-5.6-sol / -terra) is added to the UI and en/zh-CN/ko locales.

Measured: gpt-5.6-sol low/medium/high/xhigh/max/ultra (default low), gpt-5.6-luna low…max, gpt-5.5/5.4/5.4-mini/5.3-codex-spark low…xhigh.

useHarnessModels moved from ChatComposer up to ChatInterface so composer state (normalisation) and the picker share one answer.

Verification

• npx vitest run: 172/172 (new tests for label disambiguation, configured-model injection, ANTHROPIC_MODEL, Codex effort mapping, discovered-effort precedence, labelForClaudeModelId).
• npm run typecheck clean.
• End-to-end against the real Claude and Codex CLIs through the module.
• Browser: model slot sanitisation confirmed; the picker/selector UI check was cut short when the browser extension disconnected, so please give the Codex effort selector and the Claude dropdown a quick look after merging.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NBqmTvJRhyysMvMvL6cp7x

[bbsngg]

Browser verification completed after the extension reconnected (dev server running from this branch against the real database):

• On load, the corrupted claude-model (gpt-5.6-sol) was replaced by the default; the Claude dropdown showed the discovered menu plus Fable 5.1 [1M] and Fable 5.1; selecting Fable 5.1 [1M] persisted.
• Switching to Codex left claude-model untouched (claude-fable-5-1[1m]), which is the race this PR fixes.
• With gpt-5.6-sol selected, the reasoning-effort selector is present and lists Default / Low / Medium / High / Extra High / Max / Ultra, i.e. exactly what codex app-server reports. Selecting Ultra persisted and was not normalised back to Default.

[bbsngg]

Second commit (0be3cf11) addresses the duplicate entries reported after the first round ("GPT-5.6 (Sol)" next to "GPT-5.6-Sol", "Fable" next to "Fable 5", …):

• Codex built-in list now uses the ids codex-cli serves (gpt-5.6-sol replaces the retired gpt-5.6; gpt-5.4-mini, gpt-5.3-codex-spark added; default follows).
• The picker hides a built-in entry the harness has retired unless it is the selected value (then labelled "not in CLI list"). The API payload is unchanged (deprecated: true).
• Claude menu labels are expanded from the CLI description: "Default (Opus 5 with 1M context)", "Opus 5 with 1M context", "Fable 5 [1M]", "Sonnet 5", "Haiku 4.5". Descriptions show as tooltips in both pickers.

Verified in the browser on this branch: Codex dropdown shows exactly the six CLI models, no duplicates; Claude dropdown shows the labels above followed by the built-ins (Fable 5.1 [1M], Fable 5.1, Opus 5, Fable 5, Opus, Opus (Plan Mode Only), Sonnet [1M], Opus 4.8/4.7/4.6), every label distinct. 173/173 tests, typecheck clean, CI green.

[bbsngg]

Third commit (92114313): the remaining Fable duplicates ("Fable 5 [1M]" from the CLI next to "Fable 5" from the built-in table, same for 5.1) came from showing the compiled-in table alongside the CLI menu. Discovery now marks table entries builtIn: true, and once the harness has answered the picker reads like Claude Code's own /model menu: only what the CLI reports, plus the configured model and whatever is currently selected. The rest sit behind a "Show N more built-in models" row, so an explicit version such as Opus 4.6 is still one click away.

Verified in the browser: the Claude dropdown shows Default (Opus 5 with 1M context), Opus 5 with 1M context, Fable 5 [1M], Sonnet 5, Haiku 4.5, Fable 5.1 [1M] and a "Show 9 more built-in models" toggle; Codex shows its six CLI models with the retired ones folded. 173/173 tests, typecheck clean, CI green.

[bbsngg]

Fourth commit (751c125c), per review: once the harness has answered, its list is the source of truth. The "show more built-in models" toggle is gone; the picker lists only what the CLI reported plus the currently selected value (so a saved preference the CLI did not list never vanishes). The compiled-in table is now purely the fallback for when discovery is unavailable. 173/173 tests, typecheck clean, CI green.