fix(agentic-enhancer): cap conversation input to avoid context overflow

[gadievron]

The agent built the LLM conversation with no input budget: get_user_prompt
inlined primary_code verbatim (prompts.py) and the loop appended raw,
untruncated tool results every iteration via json.dumps(result) (agent.py
append sites). On a large unit the constructed input grew without bound across
iterations until it overflowed the model context (400), losing the
enhancement. MAX_ITERATIONS=20 bounds depth but not per-item size.

Operative guard: an input budget at the consumption points.

• prompts.py: cap inlined primary_code (MAX_PRIMARY_CODE_CHARS).
• agent.py: add cap_tool_result_content() and use it at both tool-result
  append sites (MAX_TOOL_RESULT_CHARS).
• agent.py: graceful anthropic.BadRequestError (400) fallback returns a neutral
  incomplete result instead of crashing the enhancement.

A deferred sibling is out of scope here: tools.py _read_function returns the
whole function body uncapped; the consumption-point cap above contains it.

Tests: tests/test_agent.py (5 tests, RED before fix). All passing.

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com

Coordination

Touches a file also modified by in-flight PR #51/#60 (region-disjoint; textual merge only).

[gadievron]

Superseded by #133 (2026-06-24), which already landed the input-cap fix on master: MAX_PROMPT_CHARS/MAX_TOOL_RESULT_CHARS + cap_tool_result_content in agent.py, MAX_PRIMARY_CODE_CHARS/_cap_primary_code in prompts.py, and tests/test_agent.py. This branch is anchored to the pre-#133 dict-shaped tool_results and now conflicts. Closing.

Note for maintainers: the only novel bit here not on master is the except anthropic.BadRequestError graceful-400 fallback — defense-in-depth behind caps that already contain the failure mode, and written against the obsolete AgentResult/dict shape. If wanted, it can be re-filed fresh against the current ToolResultBlock structure.