# Existing S2 A/H2 tail diagnostic — light review before execution Claude1794966078.809549/h-c907cdaf authorizes this descriptive-only CPU diagnostic on the already seen552 parts/204 components, under3600 CPU-seconds. It has no gate, does not reverse the S2 decision, starts no item c fit, and changes neither item1 nor its postfit barriers. Mold's full05344bac specification is normative; future c also retains that specification. Entry `tools/s2_tail_diagnostic.py run artifacts/phase2-h2-s2-tail-diagnostic-2026-11-18`, original reproduction venv, project cwd. `prepare_s2_tail_diagnostic.py` reads metadata only. Before any target computation, MeltAssist performs a light code/spec check. An executor receipt records that actual review, source receipt hash, current plan hash and source map; the entry verifies it. No five-field implementation/proof ceremony or new checkpoint review. A diagnostic-local copy of the shared CPU guard adds a running coordinator allowance; the shared guard remains unchanged. One attempt, CPU24 worker/25 coordinator, no GPU/model loading/inference or labels decoded from archives. H2 full fields are the canonical frozen S2 manifest (including exact historical S1 fold0). Full A comes from the corresponding frozen H1 `A_seconds` fields already used to form the S2 baseline. Every case verifies both raw-ID arrays and complete node counts, geometry and cached lumped mass equality, original query membership, and exact A/H2 query values against stored S2 values before reading the existing seconds.npy target once. Input hashes are checked at execution. No predictor is run. The old S2 query last-location values are copied unchanged for comparison, separately from new geometry-domain diagnostics. Each case normalizes time by original t_set, keeps negative predictions, obtains mass w from existing m_i=Σ|K|/4, and computes true/A/H2 fractional last1% volume tails. Raw-ID ordering does not cut a plateau. Stable compensated sums determine plateau mass, residual and total tail mass;10^-12 correction/volume bounds are enforced and recorded. `E_tail=(Miss+False)/2` is computed with the equivalent `1−I/.01` value and floating-point identity discrepancy also retained; min intersection ensures identical fractional tails have exactly zero directional loss. Constant truth stays in the roster with its marker. No query multiplicity enters volume measure. Peak certificates and categories follow the same full/query domains, ε, r=.10Rg, strictγ>2ε, 64u near-boundary rule, exact raw-ID peak ties, peak-set/tail-support distances and true-time drop. Positive-infiniteγ from an empty exterior is stored as null plus explicit gamma_positive_infinity/trivial_radius=true so JSON contains no invalid infinity. The strict certificate and its boundary-robust interpretation are both kept. Failure of a certificate does not imply a branch error; no branch is identified from Euclidean distance. Outputs include552 per-part JSON records, both arm tail membership threshold/volume/tie summaries, Miss/False/E_tail, query/full diagnostics, original query last scores, negative counts and target-access log. Summaries give component-equal paired H2−A differences and2000 bootstrap intervals/seed2026111408 for tail, original point and continuous diagnostics, shared draws on identical component support. Undefined values exclude the whole affected component for that metric with denominators reported; no missing component becomes a zero. Categories and certificates are reported as part counts with denominators; categories additionally have component-equal fractions. These are exploratory descriptions, not a decision rule or proof of mechanism. Full predictions are reused, not copied or rewritten. Budgets:3600 total CPU-s, including180 setup reserve (covers metadata preparation and synthetic tests), a3300-CPU-s serial worker and actual coordinator/reaped-child accounting. The diagnostic-local guard samples coordinator CPU from entry start on each live poll and stops/kills/reaps the worker at90 coordinator CPU-s, retaining the full guard failure receipt. Thus180+3300+90=3570 leaves30 CPU-s for sampled-limit overshoot, cleanup and final receipt/accounting within3600; final accounting still rejects any overspend. The coordinator measurement includes configuration, polling and guard work, not just the scientific worker. This is a sampled CPU boundary, with actual polling gap retained; no automatic retry follows a stop. The worker has7200 wall-seconds and4GiB RSS; guard output is capped at1GiB. Per-part diagnostic output is small; no full-field copies or pairwise distance matrices. A numerical/input/hash error stops the batch and preserves partial records; no silent retry or fabricated552 table. Review and publication overhead are disclosed separately from the scientific worker; they run no scientific computation. Synthetic acceptance is required before review: fractional identical/disjoint tails, constant truth, one heavy node/nonuniform weights, ID permutation, monotone transformation, duplicate query records, zero-remainder boundary, strict certificate equality, sufficient-certificate/nonconverse/empty exterior, exact other-peak and near-tail categories, plus missing/mismatched review rejection. Synthetic arrays only; no real target values before the light review. No existing training or postfit source is edited. Light review1794968673 was BLOCKED solely because the prior total CPU check occurred after exit. Its original plan30e84541 and all source bytes are retained under `before-budget-fix`, and the actual BLOCKED receipt is preserved. The local budget fix adds a synthetic real-child negative test: injected prior coordinator usage plus actual polling CPU crosses90, stops while worker usage is below3300, kills/reaps the sleeping child and verifies the failure receipt. No scientific arrays are read by this test. Only this budget change and refreshed source/plan identities require the same light review follow-up.