experiments: per-victim seeded arrival order (ADR 0044) + the deaf/mute row pair — the N-042 instrument pass

[EzequielPostan]

The instrument pass chartered at the E20 merge: the N-042 fix and the parked failure-severity run-row change, batched into one re-baseline generation per ADR 0036.

The arrival-order fix (ADR 0044, resolves N-042)

The wavefront's canonical (sender rank, addressee rank) delivery sort made every victim process its intra-wave arrivals in the same global sender-rank order, so under saturated budgets a low-ranked dialer won every admission race population-wide and a high-ranked dialer lost every one. Class-level measurands were exact under that order (ranks are independent of the class draw — every class-level prediction in E12/E18/E19 matched), but per-node tail events composed of many dial outcomes were amplified by orders of magnitude relative to the independent per-victim orders a real network approximates. Two passes measured the amplification: the E19 ordered flooder (386/400 good vs ≈ 400/400 under independent orders) and E20's seed-1148 pubseam cell (188/400 bad vs the corrected form's 84.8), both attributed to N-042 by rank dissection.

The fix: the wave sort key becomes (addressee rank, arrival key, sender rank, message identity), where the arrival key is SHA-256(lp("experiments/arrival-order/v1") ‖ run_seed ‖ lp(recipient) ‖ lp(sender)) — each recipient's intra-wave order over senders is decorrelated from every other recipient's by construction, and the whole run remains a pure function of (configuration, seeds), worker-count-independent. Driver-only; core untouched. The manifest's recorded seed-derivation rule names the arrival-key derivation.

Both frozen validation re-runs passed (the fix's own pre-registrations, recorded in the charter before the pass ran):

• Seed 1148 (pubseam-mu40-n4k-k12-bp100-cap10): 80/400 bad vs the independent-order form's 84.8 (z = −0.59), split 72 mute-class / 8 deaf-class (registered ~83 %/17 %; the deaf count's z ≈ −1.9), refusals 23 060/run vs the registered 23 063. The canonical order had measured 188/400.
• The E19 ordered flooder (ordered-flood-b50-cap16-s800): 400/400 good (canonical order: 386/400 — all fourteen strandings were the instrument's), with the class-level race columns identical to the decimal (40 617.4 rejected/run, 230 394.6 dial sends/run) — the budget refuses the same load under both orders; only the per-node race allocation changed.

The failure-severity row pair (the parked PR #152 batch partner)

Run rows gain deaf (vertices the largest component cannot reach — the in-defect) and mute (vertices that cannot reach it — the out-defect), the formal severity tables' stranded-node classification, computed in the existing condensation pass (one forward and one backward walk from the giant; a vertex disconnected in both directions counts in both classes). Raw-digraph classification under every model — M3's seed rescue shows in good, never here. The d-distribution (up_honest − largest_scc) was already derivable; the class split completes the formal folder's severity comparison surface.

The re-baseline generation (ADR 0036)

All seven baseline sweeps re-run and stored at 0764da3, compared against the 9197868 generation with the two added fields stripped: every measurand of record is value-identical — good counts equal in all seven sweeps, zero changes across topology verdicts, coverage, receipts, depth distributions, degree and standing-degree histograms, and all dial columns. The one anticipated movement is the sends/sends_by_kind/suppressed attribution triple (split-horizon origin ties among same-wave duplicate deliveries resolve by the new order): Δsends ≡ Δsuppressed on every publish slice, mean |Δ| < 1 send per publish against 10²–10⁵-send totals. docs/experiments/m2-comparison.md is re-validated and carries this generation's values (three quoted send figures moved at quoted precision; every other figure unchanged). Manifests differ in tool_commit and the extended seed_derivation only.

Records

• ADR 0044 (the instrument's arrival-order model; alternatives priced).
• N-042 resolved in specs/IMPLEMENTATION_NOTES.md with both re-runs' measured numbers.
• Addenda at the N-042 references in docs/experiments/symmetric-flooding.md, docs/experiments/m4-synthesis.md, and the program of record.
• The living row-schema contract (configs/experiments/README.md) documents the new columns and cites ADR 0044.
• Golden serialization test updated (the conscious schema-change gate); new unit tests for the classification and the canonicalisation (per-victim key order, collection-order independence).

No new registered cells: both validation re-runs reuse their frozen seeds (1148, 1137); seeds 1150+ remain untaken.

🤖 Generated with Claude Code