fix(doctor): validate providers with real test calls, not /models (#48)

[savioruz]

The doctor diagnostic warned/failed when a provider's configured model name wasn't advertised in GET /models. Proxy gateways (e.g. 9Router) that don't expose /models, or expose different model naming, produced false positives even though the provider was fully functional.

Replace the /models-list check with the same real test-call logic the setup/dashboard flows already use:

• HttpEmbedderProvider::probe() issues a real POST /embeddings and reports the returned dimension.
• HttpLlmProvider::probe() issues a real POST /chat/completions.
• doctor now reports success when the real call succeeds, and still fails on 401/5xx/unreachable/parse errors.

Regression tests use a mock with no /models endpoint at all (404) but working /embeddings and /chat/completions, covering both raw (embedder) and llm modes, plus a 500 misconfigured-provider case.

[savioruz]

Review of PR #62 at f86f287033163a673d413a522a9a424432ccfa8a (base main = b1841db, includes #41–#61 already merged), limited to origin/main...HEAD net = 4 files (doctor.rs, tests/doctor.rs, embedder.rs, llm.rs, +190/-94). Implements issue #48 only (doctor proxy false-positives). Issues #52 (MCP camelCase) and #53 (batch) are NOT part of this PR — they live in already-merged PRs on main.

🔴 Blocking

None. CI green: cargo build (49s), clippy, fmt, test (54s), audit (2m38s).

🟡 Warning #1 — LLM probe() consumes one completion per doctor run

check_provider now issues a real, minimal completion (LLM) and a real embedding call (embedder) instead of GET /models. In extract_mode = raw the LLM probe is correctly skipped, but every doctor run in llm mode burns one upstream token request. Expected for a diagnostics command (infrequent), but worth a doc note that doctor is not free — not a bug.

🟡 Warning #2 — check_models / ModelsCheck are now unused by doctor

PR #62 replaces the GET /models probe with a real test call, but memayu-llm-client still exports check_models, ModelsCheck, and per-provider check_models() methods (still pub). Nothing inside the workspace calls them now (bin/memayu doctor no longer references them). They are not dead-code errors (they're pub), so CI is clean, but they are unused public API that can silently rot. Recommend a follow-up to gate/remove them (or #[deprecated]) once external consumers are checked.

Validation

#48 — doctor probe via real test calls (net new in this PR)

• crates/memayu-llm-client/src/llm.rs: adds HttpLlmProvider::probe() — a single minimal System/User completion ("reply with exactly the JSON …") whose Ok(()) proves the endpoint is reachable, the key is accepted, and the model answers.
• crates/memayu-llm-client/src/embedder.rs: adds HttpEmbedderProvider::probe() -> Result<usize, String> — a real "dimension probe" embedding, returning the vector dim on success.
• bin/memayu/src/doctor.rs: check_provider(name, cfg, ProviderKind) replaces the old check_models(base_url, api_key, model) path; prints "reachable, key accepted, model … answered a test completion" / "… returned a {dim}-dim embedding". Local backend is unaffected (cache-dir only, no network). Result: 401/5xx/timeout → fail; success → pass. Exit 0/1 preserved.
• Targeted test: doctor_proxy_without_models_endpoint_is_healthy in bin/memayu/tests/doctor.rs — the HTTP mock has no /models endpoint at all (the 9Router/forward-proxy scenario), and doctor still exits 0 (the real test call succeeds). Plus doctor_provider_unauthorized_exits_one + doctor_server_error_exits_one for the failure side.
• Doc comment + .env.example note updated ("providers are probed with a real test call rather than GET /models").

Transitive base already reviewed (#59–#61 live in main)

• Add batch memory support and serialize MCP InitializeResult #59 batch memory (core add_memories_batch/BatchItemResult + POST /api/memories/batch + MCP memories[]), PR: Refactor setup and expose embedder backend in first-run #60 shared memayu-setup crate + web expose, feat: add /api/health readiness endpoint (#47) #61 GET /api/health, MCP serialize + camelCase tests — all APPROVED earlier; unchanged here.

Migration / breaking

• No DB schema change.
• doctor output wording changed (more accurate: "test completion"/"test embedding" vs the old "model advertised in /models"); exit codes unchanged.
• check_models/ModelsCheck still exported (see Warning Config: replace .env-only with XDG config file + first-run setup wizard #2) — no public breaking removal.

Checklist

• doctor command reports false-positive warnings for providers reached through a proxy (e.g. 9Router) #48 — doctor validates providers via real test calls, not a /models listing; proxy-without-/models scenario tested.
• local backend probe unchanged (cache-dir only).
• per-provider probe result (LLM key+model OK, embedder dims) preserved in output.
• no schema change.
• issue doctor command reports false-positive warnings for providers reached through a proxy (e.g. 9Router) #48 tagged (reference comment posted); remains open until merge.

Note: PR title has a redundant "PR:" prefix. PR description was the auto-generated commit body — edited pre-review to clarify this PR is #48-only (not #52/#53, which are already merged on main). Comment updated in place per ZeroClaw convention; no new comment.