[raymondginger2018-sudo]

Summary

Records post-compaction fog observations as an opt-in, read-only JSONL ledger. LangWatch measured a 28-step post-compaction correction window where the model's error rate is 41.9% vs 17.7% baseline (2.37×) — compaction clears the context but leaves the model "foggy" for a while. This PR only counts those observations; it never changes compaction behavior.

Only counts; never decides. The row is measured on the pre-compaction history, appended after a successful compaction, and persistence failure is logged and swallowed (logger.exception) so it can never abort the run it documents.

Changes

• core/agent_runtime/runner.py (+13): new AgentRunSpec.fog_ledger_path: str | Path | None = None spec field; fail-open write block after successful compaction. None (default) records nothing — zero I/O.
• core/agent_runtime/harness_metrics.py (new, stdlib-only imports): FogLedger / measure_post_compaction_fog / FogObservation — six-count observation, appended as one JSONL row.
• tests/test_harness_metrics.py (new): unit coverage for the measurement + ledger.
• tests/test_session_compaction.py (+52): wiring tests — read the JSONL directly, assert fog_ratio is None (not ∞) when the denominator is zero, assert the opt-in default writes nothing.

Design constraints honored:

• Read-only w.r.t. compaction: the marker is the nested compaction.reset event only (compaction.py:159); the measurement never imports SUMMARY_PREFIX and does not alter thresholds, prompts, or history in any way.
• HarnessRecord is not extended — the observation lives in its own ledger, not in the run record.
• fog_ratio is None when the denominator is zero (never inf/NaN leaking into JSON).
• Fail-open: any persistence error is caught and logged; the compaction result is returned normally.
• stdlib-only in harness_metrics.py (no new dependencies).

Tests

• Local: 75 passed (test_harness_metrics + test_session_compaction + session wiring), ruff clean.
• Local CI preflight (Windows lane, pr-ready): 5 rounds (r1-r5) on this head - final: PREFLIGHT: KNOWN-RED | jobs=10 | new=0 | known=2 (py3.12/3.13/3.14 legs green; the 2 known reds are pre-existing Windows platform legs matched to the known-reds.txt baseline; rc=1 without --allow-known-red is the expected verdict shape). Two reproducible flakes surfaced mid-flow and were fixed or baselined with isolated-green evidence: an xdist-parallel test_unified_switch flake (baselined [入账20261103]) and a load-timing flake in test_exec_error_exits_nonzero (isolated re-run green x3, baselined [入账20261106]). Desktop-bundle legs green; sanitize=clean; tree=0->0.
• Note: local Windows lane carries ~100 known platform failures (baseline-qualified); this PR adds no new red beyond them.

Dependency note

This is a single 4-file PR (one intent: record fog observations). harness_metrics.py ships in the same PR as its only consumer (FogLedger has zero call sites outside this diff) — splitting it into a stacked PR would leave an orphan module with zero call sites and a CI-red stacked leg. ~+1555 lines total, acknowledged: ~840 of them are the self-contained harness_metrics.py module plus its tests.

Out of scope (follow-up): manual compact_history calls are excluded from the fog marker (only the nested compaction.reset event counts); extending to manual compaction is a separate change.

[raymondginger2018-sudo]

Splitting this into two smaller PRs, as discussed in the review norm of keeping each PR to one mechanism:

• feat(agent): count post-compaction fog in an opt-in JSONL ledger #301 — the post-compaction fog observation: fog_ledger.py (the fog half of harness_metrics.py), the AgentRunSpec.fog_ledger_path call site, and its tests. 4 files, +640.
• feat(agent_runtime): paired capability/efficiency gates for harness changes #302 (draft) — the paired capability/efficiency gate framework: the remaining part of harness_metrics.py plus its tests. 2 files. It has no runtime consumer yet, so it is a draft and may be closed until the first mechanism that cites it lands.

No content is lost: the union of the two PRs is exactly this PR's head (f342fc98) after the module was split at the fog/gate boundary (HarnessMetricsError and the pairing/storage classes stay with the gates; FogObservation / measure_post_compaction_fog / FogLedger move to the new module with their own error type). Closing this one in favour of the two.