Bench: verify v3 memory and scaling acceptance execution #837

Follow-up of #776: X02 in the remaining exclusions and follow-ups table, finalized in #834.

Disposition: ENVELOPE / acceptance, execution unverified.

Milestone: v3.0.0 — Observed enforcement of release memory/scaling bounds is still missing; source routing alone is not release verification.

Remaining scope recorded by the audit:

Retained-state/RSS/scaling bounds, nonlinear/multistratum recursion and derived-fact exhaustion, million-fact transactions and custom chunk bounds are not Bencher memory metrics. Observe the revised v3 smoke schedule or dispatch it and inspect exact SHA/test names/bounds before claiming these enforced in nightly CI. Accounting itself has no disabled-engine control; unlimited only disables the finite cap.

C19/C24/C30 cover representative latency, not bounded retained state or peak RSS. The four C24 IDs memory_meter/{count,hash_groups}/{unlimited,limited}/10k use a checkpointed read-only 10k fixture, unlimited versus an 8 MiB cap; the 4 KiB QRY-012 preflight is untimed. Accounting remains enabled in both modes: unlimited removes the finite cap, not accounting overhead. C30 only times 65,535/65,537-fact boundary cases around the default 65,536 chunk and one late API-011 rollback; it does not establish million-fact/custom-chunk bounds.

The revised default-branch .github/workflows/smoke.yml routes main smoke at 05:00 UTC and pinned v3 acceptance at 17:00 UTC. Observe the revised schedule or dispatch it. Audit all five acceptance jobs and their actual named tests:

• recursive_acceptance_test: nonlinear/multistratum recursion, derived-fact exhaustion and scaling; bench profile, ignored tests, 45-minute budget.
• memory_limit_acceptance_test: retained state/RSS, finite-limit exhaustion/QRY-012; bench profile, ignored tests, one test thread, 30-minute budget.
• limit_acceptance_test: limit/top-k retained-state and scaling bounds; bench profile, ignored tests, 30-minute budget.
• large_transaction_acceptance_test: million-fact transactions and custom transaction_chunk_facts bounds; bench profile with test-hooks, ignored tests, 30-minute budget. The public option defaults to 65,536 and clamps to 1..=1,000,000.
• size_per_fact_in_433_shape_1m: one-million-fact storage-size bound, including ignored lib test, 15-minute budget.

The audit's last inspected smoke run 41637991836 used an earlier workflow and only main smoke. The validation benchmark run does not invoke these acceptance jobs. Why outside the latency shortlist: Bencher latency is not an RSS/state metric; these require separate assertions and acceptance evidence.

Acceptance criteria:

• Link a successful run of the revised v3 acceptance route, retaining measured source SHA, workflow SHA, exact commands/features, actual test names and complete output for all five jobs.
• Inspect and record the numeric RSS/retained-state/storage/scaling bounds and fixture sizes actually asserted; verify no ignored target silently ran zero intended tests.
• Confirm nonlinear/multistratum and derived-fact exhaustion, QRY-012, limit bounds, million-fact transactions and custom chunk bounds executed and passed; state any platform-specific measurement limitations.
• Verify the scheduled v3 source cannot silently skip acceptance, and record observed job durations against the 45/30/30/30/15-minute budgets.
• Preserve the distinction between finite-cap overhead and always-enabled accounting; do not claim an accounting-disabled control or Bencher memory enforcement.

Run heavy builds, enumeration, measurements and comparisons on idle hardware; retain reproducible evidence and clean temporary scratch afterward. This is benchmark/verification work aligned with PHILOSOPHY.md's reliability, portability and simplicity principles; any production change needs its own review.

Evidence: docs/bench-coverage-v3.md; validation run 41642356471. The validation run establishes only its recorded representative latency coverage; it does not complete this follow-up.