docs: document local model benchmark profiles

Document bounded local OpenAI-compatible provider behavior and keep
benchmark profile comparisons separate between neutral, generic-static,
and regression modes.

Expose the review toggles and bounded local model setting in the public
configuration docs.

Assisted-by: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Anders Heimer <anders.heimer@est.tech>

Author: anders-heimer

docs/benchmarking.md

@@ -60,6 +60,22 @@ The tool prints a summary to the console:
 Detailed results are written to `benchmark_results.json` in the current
 directory, including the AI judge's explanation for each finding.
 
+## Local Model Benchmark Profiles
+
+When benchmarking a bounded local OpenAI-compatible model, report these
+profiles separately:
+
+| Profile | `enable_static_bug_seeds` | `enable_targeted_bug_pattern_prescan` | Purpose |
+|---------|---------------------------|---------------------------------------|---------|
+| Neutral | `false` | `false` | Model-quality runs without regression aids. |
+| Generic static | `true` | `false` | Measures opt-in deterministic diff-local bug-pattern detectors without the LLM prescan. |
+| Regression | `true` | `true` | Tracks regression behavior with all opt-in regression aids enabled. |
+
+Static seeds are opt-in generic bug-pattern detectors. For example, the skb
+fragment seed looks for diff-local skb fragment append/growth sites that lack an
+apparent `MAX_SKB_FRAGS`-style capacity guard; it is not keyed to a specific
+benchmark commit. Keep seeded and neutral results separate in benchmark reports.
+
 ## Re-evaluating existing results
 
 If you have already run ingestion and reviews but want to re-score with

docs/configuration.md

@@ -154,6 +154,8 @@ Optional array of additional git remotes to track.
 | `email_policy_path` | string | `"email_policy.toml"` | Path to the email policy file. |
 | `max_total_tokens` | integer | `5000000` | Maximum cumulative uncached tokens (input + output) per review. Cached tokens are excluded. Set to 0 to disable. |
 | `max_total_output_tokens` | integer | `500000` | Maximum cumulative output tokens per review. Set to 0 to disable. |
+| `enable_static_bug_seeds` | bool | `false` | Enable deterministic proof-required bug-pattern seeds for local regression profiles. Not intended for neutral production review. |
+| `enable_targeted_bug_pattern_prescan` | bool | `false` | Enable the LLM-based diff prescan for selected high-signal kernel bug-pattern classes. The prescan is keyword-gated, capped, and should stay disabled for neutral production review while evaluating that mode. |
 
 ## email_policy.toml
 

docs/examples/Settings.openai-compat.toml

@@ -7,3 +7,6 @@ model = "gemma-xyz"
 base_url = "http://localhost:8080/v1"
 context_window_size = 8192
 max_tokens = 2048
+# Set true only for local/Ollama-style models that need bounded host-side
+# exploration, tool, prompt, and finalization controls.
+bounded_local_model = false

docs/llm-providers.md

@@ -258,3 +258,9 @@ cp docs/examples/Settings.openai-compat.toml Settings.toml
 ```
 
 Adjust `base_url` to point to your provider's endpoint.
+
+For local/Ollama-style models that need host-side controls, set
+`[ai.openai_compat] bounded_local_model = true`. This enables bounded
+exploration, finalization-only schemas, Todo tool filtering, prompt preflight
+margins, and minimal fallback. It is not enabled merely by choosing
+`provider = "openai-compatible"`.