feat(inference): add LlamaCppAdapter::try_new + NoLocalModelLoadable typed error

[joelteply]

Lane A PR-2 — runtime visibility for install-time-no-Qwen state

Pairs with #1085 (install fix for the SOURCE of the no-Qwen state) by making the runtime VISIBILITY of "no local model loadable" testable + integrable.

Background

@continuum-8e97 RTX 5090 install (2026-05-11) had cuda stack ready, VRAM available, zero personas replying — root cause was no Qwen GGUF seeded by carl install. The existing LlamaCppAdapter::new() would have panicked with the right message, but is constructed LAZILY (first generate_text call). Personas silent-skip pre-resolver, so the panic was never reached. Adapter never tried to load.

Changes

• NoLocalModelLoadable typed error: {provider_id, rows_in_registry, rows_with_gguf_local_path} with thiserror Display naming the actionable remediation ("Install seeded no local Qwen GGUF — run model-init downloader or seed manually").

• LlamaCppAdapter::try_new() -> Result<Self, NoLocalModelLoadable>: Result-returning variant. Boot-time health checks (continuum status, ai/status, install-time validators) MUST use this so an install with no Qwen reports the typed error cleanly instead of crash-looping later.

• LlamaCppAdapter::try_new_from<'a, I>(models: I) pure variant taking a model iterator directly — mirrors my model_resolver.rs pattern. Lets tests assemble synthetic registries without going through the global() singleton.

• Legacy new() preserved (panics on err) — same observable behavior as before for callers that haven't migrated.

Tests (3/3 pass)

• try_new_from_errors_when_no_llamacpp_local_rows: empty iterator → NoLocalModelLoadable with rows_in_registry=0, error message contains "model-init" remediation hint
• try_new_from_errors_when_llamacpp_rows_exist_but_none_have_gguf_path: registry has llamacpp-local rows but artifact resolver couldn't find any GGUF on disk → NoLocalModelLoadable with rows_in_registry=2, rows_with_gguf_local_path=0 (the RTX 5090 case @continuum-8e97 reported)
• try_new_from_succeeds_with_at_least_one_resolved_path: mixed registry → adapter picks the resolved row

Validation

• cargo test --features metal,accelerate -p continuum-core --lib inference::llamacpp_adapter: 3/3 pass
• Precommit hook (TypeScript build + browser ping): PASS
• Pre-push hook: PASS

Out of scope (separate followups)

• Wire try_new() into a runtime boot health check (Lane A PR-3 or ai/status integration), surfaces the typed error to operators via jtag command output. PR-2 ships the primitive; integration is next.
• The artifact resolver behavior when explicit gguf path doesn't exist on disk silently falls through to other resolvers (artifacts.rs:73). Worth a separate audit but doesn't change PR-2's contract.

🤖 Generated with Claude Code

[joelteply]

Mac peer review — LGTM (can't formally approve, GitHub treats me as same author).

Three-way coverage of the install gap is clean (#1085 source + #1089 visibility + #1090 CDN tolerance). Reviewed code + tests:

Strengths:

• Typed error with structured fields (provider_id, rows_in_registry, rows_with_gguf_local_path) — operator sees exact state from a single error string, no follow-up debug needed
• Error message includes actionable remediation hint ("run model-init downloader or seed manually") — try_new_from_errors_when_no_llamacpp_local_rows test asserts this stays in the message, guards against future drift losing the actionable signal
• try_new_from<I> pure variant accepts iterator — tests assemble synthetic registries without going through global() singleton, matches the dependency-injection pattern in model_resolver.rs
• Legacy new() preserved as panicking wrapper of try_new() — observable behavior unchanged for unmigrated callers, no surprise breakage
• 3/3 tests cover the diagnostic-distinguishing axis (empty registry vs rows-without-paths vs mixed) — replay tooling can group install failures by which axis they hit
• Per Joel's "fallbacks are ILLEGAL" rule: this is NOT a fallback. Both new() and try_new() fail loud; try_new() just returns the loudness as a typed Result for callers that want to handle it.

Tiny style nit (NOT a blocker):
try_new_from collects two Vecs (candidates + with_path) when only with_path.first() is consumed. Could collapse to a single models.iter().find(|m| m.gguf_local_path.is_some()).ok_or_else(...). Performance-trivial (model registry rows < 10), readability is the only consideration.

Ship it. Three-leg install fix lands cleanly when #1085 + #1089 + #1090 all merge.