v0.68.2

Greenfield calibration #10 hotfix bundle — three v0.68.1 follow-throughs. Direct codebase validation against greenfield's running state confirmed two FAILs and one PARTIAL from the calibration #9 fixes that need v2 treatment. All three are surgical: idempotent self-healing for workflow_id_history, switching the graphify probe trace-check from a fixed minutes window to the session anchor, and switching the ad-hoc cleanup cutoff from calendar age to the prior workflow's created_at. Smoke: 758 → 761 passed, 0 failed (+3 gates P1-P3).

Fixed

• workflow_id_history self-healing on every state update (H2-v2). The original H2 fix only seeded history when the array was absent — but greenfield's history was created by v0.68.0 as [current_only] (missing original), then accumulated rotations, never gaining the original_workflow_id. Worse: when init.cjs strips workflow_id + created_at and forces the first-activation branch, the NEW workflow_id wasn't appended to existing history either. Greenfield's evidence: history [995823e0, 9fa91f3a, 5ab90124, 4e954a3d, 38c12b15] was missing BOTH the original (647d32e5) AND the current (a57aa9c2). Fix moves history maintenance OUT of the conditional branches into an idempotent post-step that always ensures {original, current} ⊆ history — prepends original if missing, appends current if missing. Safe to run on every updateState call. Resolves the G6 PARTIAL (5-call gap between mcp-stats --workflow-id and --since-workflow-created) as a downstream consequence.
• memory validate graphify probe uses session anchor by default (H10-v2). The v0.68.1 H10 fix used a 5-minute window to check whether orchestrator MCP calls succeeded recently; greenfield's calibration #10 evidence showed the validate runs HOURS after the graphify burst (last call at 23:39, validate at 10:00 next day — 10h gap). No fixed minutes window works for bursty + quiet patterns. New default: read workflow.yaml::first_created_at and count successful graphify calls since session start. The semantic: "if THIS session ever successfully called graphify, the probe failure is anomalous and the warning is a false positive". memory.graphify_probe_trace_window_minutes config key still available for projects preferring a sliding window. Greenfield live check: validate now correctly downgrades to info/graphify-probe-transient ("16 graphify MCP calls succeeded since session start") instead of warning/graphify-unreachable.
• cleanupStateFiles accepts adHocCutoffMtime for explicit cutoff (H1-v2). The v0.68.1 adHocStaleDays=1 calendar-age gate was too lenient for greenfield's multi-PR-per-day pattern — 16 ad-hoc files from yesterday's session survived because they were <24h old. New adHocCutoffMtime opt (ISO timestamp string) takes precedence when set; init.cjs reads workflow.yaml::created_at BEFORE the strip+restamp and passes it as the cutoff. Anything ad-hoc with mtime older than the PRIOR workflow's start gets archived. Falls back to adHocStaleDays when created_at unavailable. Catches the multi-PR-per-day residue without breaking current-session work-in-progress.

Added

• memory.graphify_probe_trace_window_minutes config key — override the session-anchor default with a sliding window in minutes when project's call cadence prefers a strict time bound.
• 3 smoke gates P1-P3 locking each H-v2 fix to a live fixture (workflow_id_history idempotency, session-anchor trace count, adHocCutoffMtime semantics).

Changed

• recentSuccessfulGraphifyTraceCount(arg) signature widened. Backward-compatible with the legacy (minutes) call form (including 0 for explicit empty window). New ({sinceSessionAnchor: true}) opts-object form uses workflow.yaml::first_created_at as cutoff. Default (no arg) is session-anchor with 24h fallback when no workflow.yaml exists.

North-star alignment

• #1 coordination: history is now self-healing across upgrade boundaries (H2-v2); validate's verdict aligns with orchestrator's reality (H10-v2); init's ad-hoc sweep aligns with workflow-start semantics (H1-v2).
• #2 code quality: false-positive graphify-unreachable warnings stop appearing in healthy sessions (H10-v2); cross-PR ad-hoc residue clears (H1-v2).
• #3 token efficiency: validate's diagnostic prose is honest (info vs warning) — orchestrator stops allocating attention budget to false alarms (H10-v2).
• #4 3rd-party integrations: mcp-stats --workflow-id queries return correct counts after upgrade (H2-v2 + G6).