Source the Claude Code model catalog from the t3code manifest

[Tryanks]

What changes

Claude Code's model catalog is no longer hand-written in claude.rs. It is resolved from the t3code project's model-manifest.json (MIT), which lists each model's slug, name, option descriptors (effort levels with defaults, fast mode, context window), Claude CLI min/max version, and the launch adapter mappings (effortMap, [1m] suffix, context window tokens).

• crates/agent/src/claude_manifest.rs is the single owner: serde decode of the fields tcode reads, validation (version, unique slugs, profile references, semver bounds, non-empty ids), resolution to ModelSpec, and the refresh.
• A verbatim copy of the upstream file is bundled (claude_model_manifest.json) as the offline fallback. At runtime, under the process feature, the same file is fetched from raw.githubusercontent.com at most once per hour (5 min back-off after failure, 10 s timeout, 1 MiB cap) and the last good copy is cached in the data dir. A bundle newer by updatedAt outranks the cache. Invalid data never replaces a usable catalog; the network never fails list_models.
• The fetch is gated by the existing "Provider update checks" setting; with it off, only the disk cache and bundle are used.
• Removed: built_in_models, model_available, normalize_claude_cli_effort, the native context-window table and the descriptor builder helpers. Manifest effort is translated to tcode's reasoningEffort in one place; everything in the UI is unchanged.
• list_models takes a CatalogRefresh { cache_dir, network } supplied by the runtime's existing provider refresh path.

Upstream data is adopted wholesale; there is no local override table.

User-visible differences (from adopting upstream data)

• Fable 5.1 / Fable 5 default effort: high → medium.
• Opus 4.7 and 4.8 lose the Context Window selector (fixed 1M). Opus 4.5, 4.6 and 4.7 gain Fast Mode.
• Opus 4.6 default window 200k → 1M; Sonnet 5 default 1M → 200k (launches with autoCompactWindow 200000).
• Fable and Opus 5 with the default 1M selection launch as <slug>[1m]; verified with claude 2.1.272 that the CLI accepts it and reports the same 1M window.
• Fable 5.1 is the manifest default model.

Tests

Synthetic manifest tests cover: descriptor mapping (defaults, boolean, unknown descriptor types dropped), 11 invalid-manifest rejections, install keeps the previous catalog on invalid data, TTL/retry/network-off pacing including a clock moving backwards, disk cache adopted only when not older than the bundle, effortMap application (ultracode→xhigh, ultrathink→no flag, per-profile xhigh→max / max→high, passthrough), min/max version gating, [1m] suffix and autoCompactWindow semantics across window profiles.

Checks run

• cargo fmt --all --check — pass
• cargo clippy --workspace --all-targets --locked -- -D warnings — pass
• cargo build --workspace --locked — pass
• cargo test --workspace --locked — pass (agent 182)
• cargo machete — no unused dependencies
• RUSTFLAGS='-D warnings' cargo check -p tcode-web --target wasm32-unknown-unknown --locked — pass
• cargo run -p agent --example probe -- claude "Reply with exactly: PONG" /tmp/smoke — PONG, context window 1000000; with --effort ultracode the spawn args are effort=xhigh settings={"ultracode":true}.
• iOS/Android not run locally.

Docs: docs/claude-models.md describes the source and cache policy.