fix: send reasoning_effort for reasoning-tuned local models via ollama backend

[AutogenousEngine]

Fixes #2932.

Problem

The ollama backend config never set reasoning_effort at all -- only gemini did ("reasoning_effort": "low"). Benchmarking nemotron-3-super:120b against Claude/GLM-5.2 on identical corpora, it produced consistently sparser extraction, and on a larger corpus dropped a file entirely after exhausting the bisect-and-retry timeout path. No JSON parsing was involved (.content was never hollow) -- see #2932 for the full empirical writeup, including raw /v1/chat/completions measurements showing reasoning_effort: "high" roughly halves both reasoning length and completion tokens for this model relative to omitting it or "low".

Since the model's chain-of-thought narration (landing in a separate message.reasoning field on Ollama's OpenAI-compat endpoint) shares the same request's wall-clock and the client's --api-timeout, an unset reasoning_effort on a verbose local reasoning model can burn most of that budget on narration alone -- which is what pushed the real multi-file chunk into repeated bisection and eventual file loss.

Fix

Adds _resolve_reasoning_effort(cfg, model, backend), used at both call sites that previously read cfg.get("reasoning_effort") directly. Precedence:

1. GRAPHIFY_OLLAMA_REASONING_EFFORT env var (ollama only) -- explicit override; accepts "none"/"omit"/"default" to force omission even for a recognized reasoning model, mirroring GRAPHIFY_LLM_TEMPERATURE's existing convention.
2. The backend config's own default (gemini, unchanged).
3. "high" for ollama models matching the same reasoning-model families _THINK_BLOCK_RE's comment already names (nemotron, deepseek-r1, qwq) -- reusing recognition the codebase already committed to elsewhere, not inventing a new category.
4. None otherwise -- every other backend/model is unaffected, including non-reasoning models served through ollama (the backend's own default model is qwen2.5-coder:7b).

Testing

• New tests/test_ollama_reasoning_effort.py (8 tests): family matching (including a provider-prefixed nvidia/nemotron-3-super form, mirroring _model_requires_default_temperature's existing prefix-stripping), the default-for-reasoning-models / none-for-others split, non-ollama backends being unaffected by the env var, env override winning over the default (including for non-reasoning models -- an explicit choice always wins), the none/omit/default force-omit sentinels, and a backend-config default (if one is ever added for another ollama-family entry) taking precedence over the reasoning-model fallback.
• uv run pytest tests/test_ollama_reasoning_effort.py tests/test_llm_backends.py tests/test_llm_parser.py tests/test_llm_parser_reasoning.py -q -- 133 passed.
• uv run pytest tests/ -q -- same 18 pre-existing failures as an unmodified v8 checkout at this depth (shallow-clone git-history gaps in test_skillgen, environment-dependent test_ollama*/test_ollama_retry_cap failures) confirmed by running the identical command against git stash-ed clean tree before this diff; this change adds zero new failures.

Happy to adjust the default (e.g. a different effort level, or gating it behind an opt-in flag instead of a name-matched default) if you'd rather ship it more conservatively.

[graphify-labs]

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

Formal verification. No changes could be formally verified in this run.

---

Graphify review — findings

Adds _resolve_reasoning_effort (plus _ollama_model_is_reasoning) to default reasoning_effort="high" for recognised local reasoning families (nemotron, deepseek-r1, qwq) on the Ollama backend, with a GRAPHIFY_OLLAMA_REASONING_EFFORT env override and config-default precedence. Wires it into extract_files_direct and _call_llm in place of the raw cfg.get("reasoning_effort") reads, and covers the precedence rules in a new tests/test_ollama_reasoning_effort.py.

No blocking issues surfaced. 1 lower-confidence candidate did not survive cross-model review.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 756 functions depend on the 183 functions this change touches.

Health — this change adds coupling hotspots:

• new: deduplicate_entities() — 63 callers, 21 callees
• new: build_merge() — 46 callers, 14 callees
• new: extract_files_direct() — 17 callers, 21 callees
• new: extract_corpus_parallel() — 26 callers, 11 callees
• new: _call_claude_cli() — 31 callers, 9 callees
• new: dispatch_command() — 2 callers, 119 callees
• new: _extract_with_adaptive_retry() — 22 callers, 10 callees
• new: _call_llm() — 11 callers, 19 callees
• …and 15 more — each is listed as a finding

Verification — 756 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 480 function(s) in the blast radius were not formally verified this run

Formal verification

Could not verify: Could not verify \_call\_llm.

The verifier did not have enough to check \_call\_llm, so it is saying so rather than guessing. No false assurance is the whole point.

Guarantee: No guarantee either way, this is an honest abstention, not a pass.

Note: Reason: not verifiable: all 23 sampled inputs raised on both versions — the function never executed, so 'no divergence' would be vacuous (mostly TypeError — names the real obstacle, not a sampling gap)

Could not verify: Could not verify extract\_files\_direct.

The verifier did not have enough to check extract\_files\_direct, so it is saying so rather than guessing. No false assurance is the whole point.

Guarantee: No guarantee either way, this is an honest abstention, not a pass.

Note: Reason: parameter `root` is annotated `Path` — outside the synthesizable primitive/collection set

· 23 more finding(s) on lines outside this diff (see the check run).