feat(freshness-monitor): wire cycle_completion rollup into the monitor (L249)

[cipher813]

What

Wires the cycle_completion() rollup (shipped dormant in alpha-engine-lib #88, v0.44.0) into the freshness-monitor Lambda — the consumer follow-up filed against ROADMAP L249. The rollup existed but nothing called it; the monitor already builds the (ArtifactSpec, CheckResult) pairs it needs, so this is the natural home.

Why the freshness-monitor (not transparency)

The two substrate consumers are independent: transparency validates transparency_inventory.yaml rows; the freshness-monitor probes ARTIFACT_REGISTRY.yaml artifacts via check_freshness. cycle_completion aggregates check_freshness results over severity:critical rows — so it belongs here, where those results already exist.

Changes

• _serialize_cycle_verdicts() — groups the probe results by (cadence, cycle_label) and rolls each group up. The registry walk covers every cadence in one 15-min pass, so per-cadence grouping is mandatory — a single rollup over the mixed-cadence pairs would conflate the Saturday / weekday / EOD / continuous cycles into one meaningless verdict. (weekday_sf and eod_sf share a date-shaped label, so cadence is part of the group key.)
• Writes _freshness_monitor/cycle_verdict.json — one verdict per cadence-cycle, with missing/stale/probe_failed/grace_period localization lists.
• Emits CloudWatch ArtifactFreshnessCycleComplete (1.0/0.0) per cadence in AlphaEngine/Substrate. Dimensioned by Cadence only — a stable, low-cardinality set ({saturday_sf, weekday_sf, eod_sf, continuous}) a CW alarm can bind to. The per-cycle cycle_label lives in the S3 artifact, not a metric dimension (a label is high-cardinality → unalarmable + costly).
• iam-policy.json — adds cloudwatch:PutMetricData scoped to the AlphaEngine/Substrate namespace via condition. The monitor emitted no CW metrics before; deploy.sh:148 (aws iam put-role-policy) applies the grant on next deploy.

Fail-loud posture

The whole rollup block is wrapped (try/except → WARN), ordered after the primary check_results + heartbeat writes. It is secondary observability hung off the primary probe pass — a CW throttle, the IAM grant not-yet-applied, or a cycle-resolution edge can never sink the monitor's primary deliverables. Recording surface = WARN log + staleness of cycle_verdict.json. Per the CLAUDE.md no-silent-fails secondary-observability carve-out (inline comment names the swallowed mode + recording surface).

This is observational only — no new alert fires (the monitor is still in OBSERVE soak, MNEMON_FRESHNESS_MONITOR_ENABLED). Alerting on cycle incompleteness is a future step gated on the monitor exiting OBSERVE.

Verification

• 4 new tests: per-cadence verdict + WARNING-excluded; complete-when-criticals-fresh; CW metric shape (namespace + Cadence dimension + values); non-fatal rollup failure (primary artifacts still written, cycle_verdicts={}).
• freshness-monitor suite: 27 → 31 passed.

Deploy note

Needs a freshness-monitor Lambda redeploy (deploy.sh) for the IAM grant to take effect — until then the CW emit WARN-logs and cycle_verdict.json (S3, already permitted) still lands, so the degradation is graceful.

🤖 Generated with Claude Code