feat(compression): config-driven per-model threshold

[Kyzcreig]

What

Surfaces the already-wired per-model compression-threshold override as a config knob. Adds an optional compression.per_model_threshold map so a model's compression trigger fraction can be tuned in config.yaml without editing Python.

compression:
  threshold: 0.8                 # global default (unchanged)
  per_model_threshold:           # NEW — optional, fraction in (0,1]
    claude-opus-4-8: 0.85
    gpt-5.5: 0.7

Why

The per-model override path already existed (_compression_threshold_for_model in auxiliary_client.py, applied in agent_init.py) but was hardcoded — only arcee-trinity-thinking → 0.75. Threshold is a direct cost lever (higher = bigger/costlier requests, fewer summarization calls; lower = cheaper turns, more summarization passes), so it should be config-tunable per model.

Resolution order (first hit wins)

1. compression.per_model_threshold[model] — config map (new)
2. _compression_threshold_for_model(model) — built-in family default
3. compression.threshold — global config
4. 0.50 — hardcoded default

The absolute trigger stays model-aware regardless: threshold_tokens = context_length × threshold, re-derived on every /model switch in ContextCompressor.update_model.

Implementation

• New _resolve_per_model_threshold(per_model, model) helper in agent/agent_init.py: exact-match then case-insensitive; accepts (0, 1] only; ignores out-of-range / non-numeric / non-dict / bad-model input (returns None + warns) so resolution falls through safely.
• Wired into the existing threshold-resolution block; built-in default kept as fallback (config wins).

Tests

28 new tests in tests/agent/test_per_model_compression_threshold.py (exact/case-insensitive match, boundary values, out-of-range, non-numeric, numeric-string coercion, empty/non-dict map, invalid model, non-string keys). 45/45 pass with the existing arcee override suite; ruff clean.

Notes

• No behavior change unless per_model_threshold is set.
• Requires gateway restart to load (import-time).

[github-actions]

🔎 Lint report: feat/per-model-compression-threshold vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 9955 on HEAD, 9954 on base (🆕 +1)

🆕 New issues (1):

Rule	Count
unresolved-import	1

First entries

tests/agent/test_per_model_compression_threshold.py:17: [unresolved-import] unresolved-import: Cannot resolve imported module `pytest`

✅ Fixed issues: none

Unchanged: 5158 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.