[raymondginger2018-sudo]

Summary

LangWatch measured a ~28-step correction window after a context compaction where the model's error rate is 41.9% vs a 17.7% baseline (2.37x): the summary restores the context budget but not the working state, so the agent keeps re-deriving what it dropped. This records the cheap half of that finding.

Each successful compaction appends one six-count FogObservation row to the opt-in AgentRunSpec.fog_ledger_path. None (the default) does zero I/O.

Only counts; never decides. The seam is the existing compaction.reset checkpoint key (read by key, no import of the compactor), the message list is never mutated, and a failed ledger write is logged and swallowed so a measurement can never abort the run it documents.

Split out of #300, which bundled this mechanism with an unrelated paired-gate framework; #300 is closed in favour of the two smaller PRs.

Changes

• core/agent_runtime/fog_ledger.py (new, stdlib only): FogObservation (six counts, None rates with no samples, fog_ratio is None — never inf/NaN — when the baseline is zero), measure_post_compaction_fog (window after each checkpoint, retained tail skipped, orphan tool results not counted), FogLedger (append-only JSONL, O_APPEND + fsync).
• core/agent_runtime/runner.py (+13): AgentRunSpec.fog_ledger_path: str | Path | None = None; after a successful auto-compaction, measure the pre-compaction history and append one row. A failing write is logger.exception-ed and swallowed.
• tests/test_fog_ledger.py (new): counting, window cap, retained tail, checkpoint detection, error-signal, non-string tool content, rates/ratio edge cases, ledger rows.
• tests/test_session_compaction.py (+52): wiring — one row per successful compaction, nothing written when compaction is skipped, ledger failure swallowed.

Notes

• fog_ratio is None rather than infinite when the denominator has no samples, so readers never parse inf/NaN.
• Rows store raw counts only; readers recompute the ratio, so a rounding change cannot make a persisted ratio drift from its own counts.
• No new dependencies; no network, no model calls, no global state.

Tests

• Local: pytest tests/test_fog_ledger.py tests/test_session_compaction.py → 36 passed; ruff check + ruff format clean; pre-commit hooks pass.
• Related existing suites (test_agent_runner_kernel, test_compaction_memory, test_compaction_tool_call_budget) pass except one pre-existing Windows environment_context failure that reproduces identically on unmodified main.