feat(models): expand provider-native capability readers

[RaresKeY]

Summary

Expand the canonical model-capability bootstrap from #2739 with provider-native catalog detection and readers while making the general fallback strictly inventory-only. The serialized record is reduced to one versioned shape, provider identity remains separate from per-model capability, and unrecognized configured providers can still list stable model identities without inferring tasks, modalities, features, controls, or limits.

This also adds bounded normalization diagnostics to Odysseus's existing LOG_LEVEL=DEBUG toggle. The event reports canonical version, provider source, detected catalog shape, fallback status, record count, and normalized capability names; it omits model IDs and raw payload values.

Stack / Review Boundary

This PR is stacked on #2739 at d8627fcd. Because #2739's head is on a fork, this upstream PR targets dev and temporarily includes the parent commits in its comparison.

The follow-up-only review range is d8627fcd..9d4d182b, containing three commits. After #2739 lands, this branch will be rebased onto current dev so the PR comparison contains only the follow-up.

Target branch

• This PR targets dev, not main. All PRs land in dev; main is curated by the maintainer at each release.

Linked Issue

Fixes #5575

Part of #2737.

Depends on #4145 / #2739.

Type of Change

• Bug fix (non-breaking — fixes a confirmed issue)
• New feature (non-breaking — adds new behaviour)
• Refactor / cleanup (behaviour unchanged)
• Documentation only
• CI / tooling / configuration

Checklist

• I searched open issues and open PRs — this is not a duplicate.
• This PR targets dev.
• My changes are limited to provider/model catalog normalization, diagnostics, and focused regressions.
• I actually ran the app (docker compose up or uvicorn app:app) and verified the change end-to-end. The catalog normalization layer is not yet wired into runtime model discovery, so this PR is validated through native-shape and fallback fixtures rather than a user-facing runtime path.

How to Test

1. Run:

   python -m pytest -q \
     tests/test_model_capabilities.py \
     tests/test_model_capability_readers.py \
     tests/test_provider_capability_schemas.py \
     tests/test_model_capability_diagnostics.py

   Expected result: 50 tests pass.

2. Inspect the fallback regression containing type, modalities, capabilities, supported_parameters, and max_model_len; verify its canonical result remains unknown with empty features, limits, and controls.

3. Set LOG_LEVEL=DEBUG, normalize a native and a fallback fixture, and verify the log includes canonical/shape/fallback metadata but not the model ID or raw payload values.

4. Run git diff --check and Python compilation for the changed modules.

Exact stacked-head adjacent validation also passes:

• 388 provider detection, endpoint resolution, model context, Copilot, Anthropic, Ollama, Mistral, Moonshot, and reasoning tests;
• 285 model route, endpoint probe, model selection, streaming, Kimi, Cerebras, vision, and local-provider tests;
• Python compilation for the changed application, capability, reader, schema, and diagnostic-test modules.

A current-dev integration replay completed with 4,659 passing tests and 12 unrelated failures. The same 12 failures reproduced on the unchanged current-dev baseline; focused capability/diagnostic tests are green on this exact stacked branch.

Visual / UI changes

N/A — backend normalization, logging configuration, and tests only; no UI rendering changed.

• Screenshot or short clip — not applicable; no visual surface changed.
• Style match — not applicable; no visual surface changed.
• No new component patterns — not applicable; no frontend components changed.
• I am not an LLM agent submitting a bulk PR. The required issue was opened first as models: expand native capability readers with conservative fallback #5575.

Screenshots / clips

N/A — no visual changes.