verify: independent re-measurement harness for the SSA throughput claim

[akutuva21]

Harness only — no production code. Three stdlib-only scripts, written to check another lane's claim rather than to change behaviour.

What this reproduces

An independent re-measurement of agent/perf-engine's SSA claim (PR #44), built from source in a separate worktree with a matched binary pair differing in one TU, pinned by blob SHA (57e504a = origin/main 6889fba vs 07baa1a = aab2c0f; git diff --stat = 1 file, +9/-93).

	claimed	measured	verdict
throughput (species=150, R=450, 2M-event cap)	1.599x	2.061x min / 2.044x median raw; 2.129x / 2.120x probe-corrected	reproduced, conservative
bit-identical trajectories	yes	22/26 models byte-identical	reproduced
trajectory physics correct	—	exact integer conservation, drift 0.0 over 300,000 events	reproduced
depOffset[idx+1] OOB write (orchPerf)	flagged	2711 .net files, 792,960 reaction lines, zero out-of-range refs	not reproduced

Arms do not overlap (base max 910k < changed min 1.79M events/sec) across 9 interleaved reps with per-round order flip. Peak child RSS identical at 21,463,040 kB.

Limits, stated rather than rounded up

• Measured at one operating point (R=450). This does not replace perfEngine's scaling table and says nothing at R=45 or R=1350.
• Identity: 2 models INCONCLUSIVE — three identical runs of the baseline binary produce different hashes, so no comparison can adjudicate them. Reported as neither win nor defect. 1 excluded (exceeded a 240s cap, no artifacts in either arm). 1 not-a-case (stale dir, replaced by test_MM/test_sat covering the same isTotalRate path, both identical).
• The identity gate is a test of the trajectory, not the engine, and has a half-life: if the SSA seed derivation changes (as fix(batch-ssa): restore the missing batchTrajectorySeed declaration — main does not compile without it #69 did), these hashes become vacuous rather than false. The conservation check is the absolute instrument and does not share that bound.
• perfBatch's PR perf(batch-ssa): compile the model once for the whole CPU pool #58 was partially verified: −1.80% vs −2.0% claimed on the small-batch config; the headline −36.4% and RSS halving were not measured here (three attempts exceeded 3 minutes each; killed rather than report a partial figure).

Measurement context

R=48 runnable, loadavg band 117–147 (coarse band only). Three declared co-tenants during the timed window, not two: my A/B arms, perfEngine's NFsim baseline, and my own identity sweep — an omission I reported and corrected. Nothing load-bearing here is a configure, ctest, or build-system gate, which is why these results are unaffected by the duplicate-target defect or the FetchContent question.

The scripts

• verify/identity_check.py — fixed-seed simulate_ssa over 25 models plus synthetic edge cases targeting the change's risk surface: negative rate constant (non-monotone fallback), zero-rate plateaus, identical reactants, a species on both sides of the fired reaction, larger reaction counts, and saturating rates where propensity reads no species at all.
• verify/compare_identity.py — separates IDENTICAL / DIFFERS / INCONCLUSIVE.
• verify/ab_bench.py — interleaved A/B, child CPU time, paired 2-event probe, per-round order flip, and a .gdat hash guard that fails the run if any arm or rep diverges.

No Python package is imported by any of them, so the editable-install shadowing and the conftest sys.path filter do not apply.

🤖 Generated with Claude Code

[akutuva21]

Superseded by reviewed PR #92, merged as 7107b6c. All three identity/A-B harness paths are carried as reviewed updated files, including fresh usable artifacts, complete sampling grids, comparable manifests and exact nonempty case coverage with regressions. Old trajectory outputs do not qualify under the repaired contract. PR92 head6d835c3 finished40successful/5skipped checks. Original author branch is preserved.