fix: generalize local-model validation bypass beyond Ollama tags

[Aneesh-382005]

Summary

• is_local_openai_model_syntax() only bypassed strict provider/model validation when the model name contained a . or : (Ollama-tag-shaped). Bare names LM Studio/llama.cpp commonly use (e.g. mistral) were rejected unless prefixed with local/.
• Fix: reuse the existing is_local_openai_compatible_base_url() check (already used to waive the API-key requirement) to also waive the model-name shape check when OPENAI_BASE_URL is loopback/private. Remote gateways keep the old heuristic, so a mistyped bare name can't silently route to a paid API.
• Scoped narrowly on purpose see Future plans.

Anti-slop triage

• Classification: actionable-fix
• Evidence: OPENAI_BASE_URL=http://127.0.0.1:1234/v1 claw --model "mistral" prompt "hi" fails today with invalid model syntax despite the server serving that model. Verified fixed against a local mock server; verified remote gateways (openrouter.ai) still reject bare names.
• Non-destructive review result: merge candidate

Verification

• scripts/fmt.sh --check, cargo test --workspace all pass, incl. 4 new tests.
• git diff --check passes.
• No secrets/unrelated churn; diff scoped to 6 files.
• Note: cargo clippy --workspace -- -D warnings fails, but identically on unmodified main (confirmed via git stash) -> pre-existing, not touched by this PR.

Resolution gate

• Follow-up to my earlier PR Feat/ollama provider #3213 (native OLLAMA_HOST support); ROADMAP entry added.
• N/A - intended to merge.
• N/A - no remote PR/issue automation actions taken.

Future plans

Deferring two follow-ups raised in #3213's review to separate PRs: named per-backend env vars (LMSTUDIO_HOST, etc.) and a config-based local-provider registry. Keeping this one small and reviewable.

[1716775457damn]

Good fix — the restriction that only Ollama-style tags (with . or :) could bypass validation was a real pain point for LM Studio/llama.cpp users with bare model names. Reusing the existing is_local_openai_compatible_base_url check to also waive the model-name shape constraint is the right approach — it's scoped to loopback/private URLs, so remote gateways keep validation. The anti-slop triage section with the concrete reproducer is appreciated.

[Aneesh-382005]

Good fix — the restriction that only Ollama-style tags (with . or :) could bypass validation was a real pain point for LM Studio/llama.cpp users with bare model names. Reusing the existing is_local_openai_compatible_base_url check to also waive the model-name shape constraint is the right approach — it's scoped to loopback/private URLs, so remote gateways keep validation. The anti-slop triage section with the concrete reproducer is appreciated.

Thanks for the review! Once this merges, I'm planning a couple of follow-up PRs for named per-backend environment variables and a config-based local provider registry. Do you think it's better to wait for this to merge before opening those, or would it be fine to work on them in parallel?