fix(plugin): remember unsupported reasoning-effort routes per bridge

[asyade]

Problem

prepareAuxiliaryCall in the DeepSeek Harness adapter probes DSH'"'"'s branded off reasoning effort before every auxiliary LLM call, and falls back to an effort-less preparation when the exact model rejects it. That fallback is correct — but for models that advertise no reasoning efforts at all (e.g. a hand-declared OpenAI-completions entry), the probe fails every time, so each summarize / retrieval-filter call pays one doomed prepareCall, one logged UNSUPPORTED_REASONING_EFFORT error row, and a retry before doing any real work.

Observed in production while running the plugin (2.0.16) against DeepSeek Harness with OpenCode'"'"'s x-preview-f-free: the harness-side api log accumulated 1,220 failed capture.summarize preparations vs 191 successes, and 1,267 failed retrieval filters vs 106 ok — every success reached only through the fallback retry, doubling round trips and flooding the log.

Fix

Each bridge now remembers the routes that rejected the probe (keyed by provider/model) and skips the probe on later calls:

• first call for a route: probe off → on UNSUPPORTED_REASONING_EFFORT, remember the route, prepare without effort (unchanged behavior);
• later calls for a remembered route: go straight to the effort-less preparation;
• models that do support off are unaffected — they keep getting the probe;
• non-unsupported errors still propagate untouched and do not poison the cache.

The memory is scoped per bridge instance (closure state, not module-global), so concurrent bridges stay independent and unit tests remain hermetic across cases.

Tests

• new: "remembers an unsupported-effort route and probes it only once per bridge" — asserts the exact preparation sequence [off, plain, plain] across two completes;
• new: "keeps remembered effort routes scoped to a single bridge instance" — guards against reintroducing module-global state;
• existing suite passes unchanged (20/20 in tests/unit/adapters/deepseek-harness-host-llm.test.ts);
• tsc -p tsconfig.json --noEmit clean.

[Memtensor-AI]

🤖 Open Code Review

Target: PR #2285
Task: 392e613ed310c57b
Base: main
Head: fix/host-llm-effort-probe-cache

✅ OpenCodeReview: Review complete: 0 finding(s) across 1 selected item(s).

Generated by cloud-assistant via Open Code Review.

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (20/20 executed). memos_local_plugin/unit: 20/20. Duration: 3s [advisory, non-gating] AI-generated tests on branch test/auto-gen-392e613ed310c57b-20260826190007: 41/43 passed, 2 failed — these do NOT affect the PR verdict; review the branch manually.

Branch: fix/host-llm-effort-probe-cache

[asyade]

Closing at the author's request — withdrawn for now.