give the OpenCode reviewer a model dropdown and hide reviewers this machine can't run

[atomantic]

Summary

• OpenCode's Model cell is a dropdown, not a text box. Its options come from the seeded OpenCode Zen CLI/TUI records, whose namespaced opencode/* ids the Harnesses page refreshes from opencode models. The opencode-<local-backend> wrappers stay excluded — their ids resolve only under the OPENCODE_CONFIG_CONTENT a PortOS-spawned provider injects, and the reviewer runs a bare opencode against the user's own config. Still free-text underneath, so a config-declared namespace can be typed.
• Each reviewer's options now union every provider record fronting its binary instead of reading one. Reading a single record is what showed the Claude reviewer the retired claude-sonnet-4-6 from claude-code while claude-code-tui already listed claude-sonnet-5. The CLI record still owns the shown default, since the reviewer is spawned non-interactively. Bedrock records are excluded (their us.anthropic.* ids resolve only in that environment).
• The Add row hides reviewers this machine can't run — CLI missing, or every backing provider record switched off — behind a +N unavailable toggle. Hidden, not dropped: both checks are local-machine-only and the reviewer list is federation-wide config, so the toggle reveals them badged with which signal fired, and an already-selected reviewer always renders its row.
• Migration 337 repairs the stale data. claude has no models subcommand, so nothing in the app can refresh a stale Claude record. It inserts claude-sonnet-5 after a retired claude-sonnet-4-6 even on a curated list, which migration 153's exact-match policy skipped.

Why the migration is additive

153/206 rewrote models only on a list matching the prior seed exactly, leaving a user who had appended one id stuck on the retired tier. 337 targets exactly those records, so it takes the opposite policy: claude-sonnet-4-6 and every tier pointer survive untouched. That id still resolves for the CLI, and the defect is the new tier being absent, not the old one being present. A record already listing claude-sonnet-5 is untouched, so it is a no-op on a seeded install and on a second run.

Test plan

• client: 1952 tests pass across components/cos, hooks, components/settings.
• New hook tests pin the union (Claude CLI+TUI, grok CLI+TUI), OpenCode's Zen-only sourcing, that a local-runtime wrapper's ids are never offered, OpenCode's deliberately-null default, and providerDisabled separating "all switched off" from "no records" and "fetch failed".
• New picker tests pin the hide/reveal behavior, both badge reasons, that a revealed reviewer still adds normally, that an already-selected unavailable reviewer stays visible, and that nothing is hidden when neither signal was fetched.
• New migration tests cover a curated list, each Bedrock record's region-qualified spelling, idempotency across two runs, non-target records left alone, and missing/malformed/non-array inputs.
• Migration verified against the four seeded Claude ids; scripts/migrations + reviewerConfig + codeReview suites pass (2072 tests).

Follow-up filed as #6072: makeSeededProviderTierMigration exists to generalize this shape but no migration calls it, so 337 is the third hand-rolled copy of the same scaffolding.