agent: prompts embed the philosophy as an invalid JSON fragment once it outgrows the caps

[phamngocquy]

Summary

All three scored prompt sites that embed the inferred philosophy hard-slice its serialization:

• agent/planner.py (plan_next_actions): json.dumps(philosophy, indent=1)[:4000]
• agent/decider.py (direction lens): json.dumps(philosophy, indent=1)[:3000]
• agent/decider.py (final synthesis): json.dumps(philosophy, indent=1)[:3000]

infer_philosophy puts the unbounded evidence list last in the serialization, so whenever the philosophy outgrows a cap the slice cuts mid-string inside evidence — the model reads an unterminated JSON fragment with the tail destroyed, in the same prompts that tell it "Respond ONLY with JSON". Real inferred philosophies do exceed these caps: the instrumented run published in #1706 measured one at 4,237 chars (evidence 2,508 chars / 13 entries) — over all three budgets. #1706 covers the judge-side sibling of this defect (benchmark/judge.py::_render); this issue is the agent-internal side, where a fix is plain prompt-assembly correctness with no scoring-surface contact.

Proposed fix: a field-aware bounding helper in agent/philosophy.py — byte-identical serialization whenever it fits the existing cap; otherwise drop whole trailing evidence entries (they are real, individually intact signals) so the prompt stays valid JSON and summary/values/merge_bar/direction survive; keep the existing hard slice only as the last resort for shapes that cannot fit even with evidence empty, so no input ever renders worse than today. The call-site budgets (3000/4000) stay unchanged, and the philosophy solve() returns is untouched — only what the agent shows itself in prompts changes.

Fixes #1962