feat(config): expose dspy.RLM max_output_chars as a budget knob

[popuku-dog]

What & why

dspy.RLM head+tail-truncates each REPL output before it enters the planner
prompt — the planner never sees the omitted middle. The cap is max_output_chars
(characters, distinct from max_tokens), but the kit only forwards
max_iterations / max_llm_calls, so it's pinned at dspy's 10000 default with
no config path.

This adds RLMConfig.max_output_chars (env RLM_MAX_OUTPUT_CHARS, default
10000 = dspy's own → no behaviour change unless set) and passes it through the
existing best-effort budget dict in _build_rlm.

Observed in a real run

When a REPL step prints more than the cap, the planner receives only the head
and tail — here it lost 25,141 of 35,141 chars mid-output, with no way to raise
the cap from config. Langfuse trace (what the model actually received):

=== Step 3 ===
Reasoning: ...
Code:
```python
"some python code"

Output (35,141 chars):

... (25,141 characters omitted) ...

[[ ## iteration ## ]]
4/10

## Checklist

- [x] `uv run --group dev python -m pytest` is green
- [x] `uvx ruff check .` is clean
- [x] New behavior has a test
- [x] No new top-level `dspy` import in a dspy-free module (`config` / `_retry` / `sandbox` / `tools` / `trace` / `skills` / `replay` / `dataset`)
- [x] The `rlm-kit/trace/v1` trace is unchanged, or the change is additive (a new optional field) — `tests/test_contract.py` still green
- [x] The public surface stays vendor-neutral (no specific downstream project names or values)