fix(compiler): don't cap concepts-plan at max_tokens=2048 (empty plans on reasoning models)

[cnndabbler]

Problem

The concepts-plan LLM call is hard-capped at max_tokens=2048:

], "concepts-plan", max_tokens=2048, response_format=_JSON_RESPONSE_FORMAT)

With reasoning / thinking models (e.g. ollama/qwen3.6:27b, deepseek reasoners), the model can spend the entire 2048-token budget on reasoning before emitting any JSON, so the response comes back empty. The plan then fails to parse and zero concept pages are generated — silently, while the doc is still marked [OK]:

openkb.agent.compiler WARNING: Failed to parse concepts plan: Expecting value: line 1 column 1 (char 0). Raw output (first 500 chars): ''
[WARN] concepts plan unparseable for <doc> — no concept pages generated.

This is the failure in #80 (model: ollama/qwen3.6:27b). Observed hitting ~50% of docs with a local thinking model before the cap was lifted.

Fix

Remove the max_tokens=2048 cap so concepts-plan is uncapped, matching the existing uncapped summary call in the same pipeline. Content-rich plans (and reasoning-model preludes) then complete instead of truncating to empty.

Note / alternative

If a hard ceiling is preferred for cost control, an alternative would be to make it configurable (e.g. a concepts_plan_max_tokens config key, default high or None) rather than a fixed 2048. Happy to adjust to whichever you prefer — but the current fixed 2048 is too low for reasoning models and fails closed (empty output → no concepts).

Fixes #80.