fix(models): bound provider-facing observation output

[snvtac]

Summary

• Add a shared observation utility that bounds provider-facing command output while preserving the original extra.raw_output for local trajectory/debug use.
• Apply the bound before rendering observations for tool calls, Responses API tool calls, and text-based observations.
• Add regression coverage for large stdout, retry-amplified provider payloads, the real mini.yaml observation template, and provider message preparation stripping extra.raw_output.

Fixes #873

Validation

• uv run pytest tests/models/test_observation_truncation.py
• uv run pytest tests/models -n auto -> 140 passed, 1 warning
• uv run pytest -n auto -> 536 passed, 76 skipped, 1 warning
• uv run ruff check src tests
• uv run ruff format --check src tests
• uv run pylint src/minisweagent --errors-only
• git diff --check

Notes

Validation used a repo-local .venv and .uv-cache; both were removed after verification along with ignored test/lint caches.

[klieret]

But we can already truncate observations with the jinja templates, why is this needed?

[klieret]

see the .yaml configs