fix(compression): honor custom provider context lengths

[Bartok9]

Problem

Compression models configured with provider: auto and an empty model inherit the main model name, but the startup feasibility check did not inherit the custom_providers[].models.<model>.context_length metadata for that model. Users with custom OpenAI-compatible providers could see Hermes auto-lower the compression threshold to the 256K fallback even when config.yaml declared a larger context window.

Closes #21947

Root cause

get_model_context_length() already supports custom_providers, and the main compressor path already resolves custom-provider context into model.context_length before construction. The auxiliary compression feasibility probe, however, only passed auxiliary.compression.context_length and did not have access to the resolved custom provider list.

Fix

Cache the resolved custom provider metadata on the agent and pass it into _check_compression_model_feasibility() when resolving the inherited auxiliary compression model's context length.

Testing

• python3 -m py_compile run_agent.py tests/run_agent/test_compression_feasibility.py: clean
• python3 -m ruff check tests/run_agent/test_compression_feasibility.py --select F: clean
• python3 -m pytest -o addopts='' tests/run_agent/test_compression_feasibility.py -q: 17 passed

Note: full-file ruff --select E,W,F on run_agent.py still reports pre-existing unrelated issues; the repo's CI uses the ruff diff workflow for PRs.

[Bartok9]

The test CI failure is pre-existing on main — verified at main SHA 1997b3ba. Not introduced by this PR.

[Bartok9]

The test CI failure is pre-existing on main — verified at main SHA 524cbabd. Not introduced by this PR.

[Bartok9]

CI note: The failing checks (test, e2e, Windows footguns) are pre-existing failures on main at 3c23b15f815e — they fail without this PR's changes. Running pytest locally against main directly reproduces the same 12 failures. None of the failing tests touch the code changed in this PR. This PR is CI-clean with respect to its own scope.

[Bartok9]

Rebased onto current main.

What I did:

• Cherry-picked the 4 fix commits (e74b5f7, 88ebe31, 5e3f521, 1593ef5) onto a fresh branch off upstream/main
• Dropped the merge commit (e8c12fdf) — main has moved on, the merge artifact was no longer relevant
• Resolved two small conflicts:
  • tests/run_agent/test_compression_feasibility.py: agent._custom_providers = [] (accepted main's value — [] is correct, None was my older draft)
  • run_agent.py: deduped a self._custom_providers = _custom_providers assignment (main had it once with a long comment; the fix commit's intent was the one-liner version which is what landed)
• Also caught + fixed 3 duplicate custom_providers= kwargs that the cherry-picks produced (test assertions had custom_providers=None, provider=..., custom_providers=[] after the merge — both kwargs collapsed correctly)

Verified: ast.parse clean on all 3 touched files (run_agent.py, agent/context_compressor.py, tests/run_agent/test_compression_feasibility.py).

Final diff: 2 files, +39 / -7

Closes #21947 (custom provider context length feasibility check).

[Bartok9]

Closing to stay within contributor PR limit. Will resubmit with fresh rebase if the issue remains open in main.