feat(sandbox): surface agent transcript errors

[iamdhrv]

Change

• persist tool status and model response metadata in Forge's existing
  .forge/history/<task>.json artifact
• inspect the last three assistant turns plus intervening tool results after a
  non-zero container exit
• surface explicit tool errors, token-limit stops, and assistant refusals at
  error level
• always log the task-specific transcript path when history is available

Why

Container stderr often contains only startup noise. Forge already persists the
Deep Agents conversation, but the host runner did not consult it and omitted
the stop/status metadata needed to identify failures. The new parser uses that
existing task-scoped JSON artifact, bounds logged excerpts to 1,000 characters,
and leaves startup unaffected when history is missing or malformed.

Checks

• uv run pytest -q tests/unit/sandbox/test_transcript_errors.py tests/unit/sandbox/test_runner_review_polling.py (31 passed)
• uv run ruff check src/forge/sandbox/runner.py containers/entrypoint.py tests/unit/sandbox/test_transcript_errors.py
• Python byte-compilation for both changed runtime modules
• git diff --check

Fixes #79

[eshulman2]

Rebased onto current main and reviewed the conflict resolution and resulting diff. The container-driver refactor is preserved, transcript extraction is wired to ExecResult, and focused tests plus Ruff checks pass. No blocking findings.