Resolves #15014

Related: #14687 — the measurement PROGRAM stays open for the operator evidence (the institution-day runs on named hardware + the hosting-bill read); this PR delivers its INSTRUMENT leaf, converged in review as the honest close-target split.

The [UNMEASURED] discipline gets its instrument: a steady-state serving-cost meter for the always-on inference load — the measurement half of "cost claims are invalid until a named measurement exists." What one institution-day actually consumes, split honestly into idle vs active phases, on named reference hardware.

The shape (entrypoint CLI + pure core, per the C1 discipline):

• ai/scripts/benchmark/helpers/servingCostCore.mjs — the PURE transforms (no Neo/AiConfig import, no clock reads — callers inject timestamps, so identical inputs are identical outputs forever): threshold-heuristic phase classification (inclusive boundary, fail-closed on garbage), window aggregation where coverage gaps are excluded and reported, never guessed into idle, and metric-bag shaping where every figure is BORN business-schema-valid — claimClass: 'measured', the exact re-run command as its falsifyingQuery, the heuristic + coverage counts as its confoundDisclaimer, deterministic idempotent METRIC identity (recomputation lands on the same node).
• ai/scripts/benchmark/serving-cost-meter.mjs — the entrypoint CLI (the sanctioned home for the AiConfig read, beside its sibling probes): resolves the sampled ports from the config SSOT's own endpoint leaves (openAiCompatible.host / ollama.host / the chroma port), re-resolves owning PIDs every tick (a mid-window server restart is sampled, not lost), ps-samples rss + pcpu, and writes a provenance-stamped JSON report (host descriptor, threshold, interval, per-role aggregates, the schema-valid bags). A vanished endpoint owner is a gap by omission — never a fabricated zero-load sample. Bags flow to the existing tenant-ingestion path; this instrument never writes the graph.
• learn/agentos/measurements/serving-cost.md — the results doc SKELETON: method + run ledger + provenance discipline, with every figure slot [UNMEASURED] by construction (zero placeholder numbers; the doc refuses them the way the schema refuses an unanchored metric). Follows the sibling measurement-doc precedent in the same directory.

What the meter deliberately does NOT do (declared in its own JSDoc): request-level token throughput (no provider-metrics dependency in v1), per-model attribution when chat + embedding share one server process (one honest role per port, never a fabricated split), anything about pricing (public substrate carries method + raw measurements only — derivations are private-side, the standing rule).

Live-proven, and two real bugs died before the formal review: (1) my own smoke caught a double-count — two endpoint entries sharing one role name made the aggregation loop emit duplicate bags, with a latent same-timestamp chronology trap had both servers been live; fixed (unique per-endpoint roles; aggregation iterates the sample streams). (2) The reviewer's pre-review live falsifier caught a client-misattribution in the sampler: a bare lsof -ti :port matches every process with a socket on the port — INCLUDING connected clients — so a 100%-CPU summarizer client was counted as ~98.7% "vector-store" load. Fixed at the root: -sTCP:LISTEN scopes ownership to the actual server (documented in the sampler's JSDoc), plus two adjacent hardenings from the same cycle — remote endpoints are skipped with a DECLARED reason (in stdout and in the report artifact — a host that isn't local isn't ours to sample, never a silent hole) and startup validation fails loud on a NaN interval/threshold before the first sample rather than hours later as a garbage window.

Evidence: L2 (the pure core 4/4 spec-pinned against the REAL businessSchema gate — verdict-object validation, determinism proof, provenance-throws) + L3 (two live smoke runs on this machine's actual resident processes; the second at the fixed shape: 2 roles, 8 schema-valid bags, report artifact written) → L3 achieved for the INSTRUMENT. Residual (L4, operator-executed by nature — declared at claim time, the evidence-ladder handoff pattern): the N-hour institution-day runs on named reference hardware (AC 1–2) and the hosting-bill console read (AC 3) — the CLI is their instrument, not their substitute.

Deltas from ticket

• The ticket's cited instrumentation pattern (the fresh-install phase-log vocabulary) did not survive into the shipped tree (its measurement anchor was retired since) — the meter follows the LIVE sibling precedent in ai/scripts/benchmark/ instead (the probe-CLI shape with the what-this-does-NOT-prove honesty sections). Recorded at intake.
• Token-throughput sampling is explicitly v1-out (no provider /metrics dependency) — a server that exposes one feeds a later leaf; the ticket's duty-cycle/memory/cpu core is fully served.
• METRIC emission ships as validated bags consumed by the existing tenant-ingestion path (the epic's own named-reuse AC), not as direct graph writes from the meter.

Test Evidence

At head 9069e7784 (post-convergence: listener-only sampling + declared skips + startup validation):

npm run test-unit -- test/playwright/unit/ai/scripts/servingCostCore.spec.mjs --workers=1
8 passed — the pure core (phase heuristic incl. boundary + fail-closed; gap-excluding aggregation with the determinism proof; schema-valid bags against businessSchema's own gate; deterministic idempotent identity; provenance-throws) + the CLI resolution helpers (window parsing fail-closed; port extraction; the endpoint-LOCALITY sampling gate; role/port dedupe with DECLARED remote skips)

node --check ai/scripts/benchmark/serving-cost-meter.mjs — passed
node ai/scripts/benchmark/serving-cost-meter.mjs --hardware dev-smoke-<host> --window 12s --interval 3
→ listener-only sampling: 2 live roles, 8 schema-valid metric bags, report artifact (with the skipped[] declarations) written under ~/.neo-ai-data/serving-cost/

Post-Merge Validation

• Operator run 1 (AC 1): one institution-day window (--window 24h --hardware <named-slug>) on the reference machine during normal operation — the report artifact fills the results doc's first ledger row.
• Operator run 2 (AC 2): the same window on a second available hardware option (availability documented; no modeled third options).
• Operator read (AC 3): the live hosting bill from the console, recorded in the ledger with its date.
• The emitted bags ride the tenant-ingestion path into the graph once the first real run exists (no smoke-run figures ingest — dev-smoke reports are instrument evidence, not measurements).

Process note: authored during the operator-granted temporary Fable 5 window.

Authored by Grace (Claude Fable 5, Claude Code). Session ef6b9a4a-54ec-4afb-8438-f89a3ee46ad2