fix(context): use Ollama loaded runtime allocation

[RaresKeY]

Summary

Use Ollama's /api/ps loaded-model allocation for context budgeting instead
of promoting a static model maximum to effective runtime truth. The runtime
response is normalized through the provider adapter layer, and
src/model_context.py remains the consumer that performs the local probe,
handles caching, and returns the effective value.

Manually added local Ollama endpoints stored as endpoint_kind="api" now probe
before the existing API/proxy short-circuit. A valid Ollama response with no
matching loaded model remains runtime-unknown, so the static model table is not
sent back as options.num_ctx. Remote endpoints and configured proxies do not
receive the probe, and non-Ollama local endpoints keep their existing request
sequence.

Stack

This draft depends on #5576 and the focused runtime-context evidence adapter in
#5628 / #5630. After those layers land, rebase this branch so the PR contains
only src/model_context.py and tests/test_model_context.py.

Related implementation

PR #5355 targets the same bug but reads /api/show model metadata as the
serving window. This implementation instead uses /api/ps, which reports the
allocation for currently running models, and keeps model maximum, configured
num_ctx, and loaded allocation distinct.

Target branch

• This PR targets dev, not main.

Linked Issue

Fixes #5193

Part of #2737

Type of Change

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

Checklist

• I searched open issues and open PRs — fix(context): probe Ollama /api/show serving window (#5193) #5355 is related but uses a different and incomplete evidence source.
• This PR targets dev.
• My changes are limited to the scope described above — no unrelated refactors or whitespace changes are mixed in.
• I ran the app end-to-end. Live Ollama validation remains to be performed; focused, adjacent, and full automated checks were run in isolation.

How to Test

1. Run:

   python -m pytest -q \
     tests/test_model_context.py \
     tests/test_model_capability_readers.py \
     tests/test_llama_server_models_url.py \
     tests/test_budget_auto_sentinel.py \
     tests/test_llm_core_ollama.py

   Expected result: 94 tests pass.

2. Configure a local Ollama endpoint as endpoint_kind="api", return a
   matching /api/ps row with context_length: 65536, and verify
   get_context_length() returns 65536 without querying /v1/models.

3. Return a valid /api/ps response before the model is loaded and verify the
   result remains unknown; load the model at 32768 and verify the next lookup
   returns 32768 rather than a cached static-table value.

4. Verify remote endpoints, configured proxies, and an unconfigured
   localhost:8080 OpenAI-compatible endpoint do not receive /api/ps probes.

Broader context, compaction, capability, Ollama, and endpoint tests pass 285
tests. Python compilation passes for all changed modules. The full suite
reaches 4,711 passing tests and 5 skips; its 12 unrelated failures reproduce
unchanged on the #5576 base.

Optional live check:

1. Start Ollama with a model loaded below its theoretical maximum.
2. Compare GET /api/ps with Odysseus's logged context length.
3. Confirm agent prompt budgeting follows the loaded allocation after the
   model is running.

Visual / UI changes

N/A — backend context discovery and tests only; no UI rendering changed.