[Clear1nsight]

Change and mathematical scope

Complete frozen precision qualification recomputes the actual ideal/error outputs after both full contracts have already produced them. This opt-in package retains those complete first-pass outputs separately inside each independently executed backend, then performs the same full comparisons. Across six fresh serial processes, median wall time falls 61.007 → 39.015 seconds (36.05% less) and median CPU falls 60.950 → 38.880 seconds (36.21% less). All observations are retained in B,C,C,B,B,C order.

Both complete Fraction and exact-integer contracts still execute all 25,984 events and 799 coordinates. Full original results, actual ideal metadata/local weights, complete error traces and every signed witness are compared. All signed final transpose checks, literal attaining fixtures, six original semantic rejections, original 120-program/4,240-event differential and three integer rejections remain. No backend, frozen source, expected receipt, timeout, arithmetic or sampled subset is changed.

Each observation binds the whole word, exact weights and rounded operation set, checks input stability, deep-copies outputs and is consumed once after its complete contract succeeds. The two backends share no new answer cache; nothing persists across commands. The Fraction module-factory seam is scoped and restored in finally. PROOF.md and COVERAGE.md explain why eliminating duplicate deterministic same-input calls preserves the comparisons. This is verification engineering: precision/physical claims, all five framework assumptions, kappa and selected metadata retain their status.

Task and overlap

Explicit donor assignment; refined replay-runtime, scope PR5831/frozen-7c5eb98/complete-single-pass-qualification. Pre-execution claim and profile/finite/first-pair progress precede expansion. Canonical main freshly fetched at entry and submission is de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8. Source PR #5831 remains 7c5eb98e7439d778141d63e801e53d4cd7e1d587, retaining PR #5826 58f93cf41751920f3d6766a39caccc2974ce18b0.

PR #5839 validates production replay on Linux and separately records the slower expanded qualification. This contribution addresses that expanded qualification. PR #5818 and PR #5832 address different literal-interpreter qualifications. Targeted PR/body searches, source PR comments and Discussion #2750 found no matching successor. This is bounded overlap coverage, not exhaustive or unpublished-work assurance.

Profiling and finite feasibility preceded construction. The nominated shared-helper group passes all ten tests after authenticating its complete fixture/history closure, but setup is only 0.287 of 51.317 seconds. Even removing all setup cannot meet 15%, so no shared helper, test, fixture, wait or platform skip is changed. The measured precision duplication justified the one candidate.

The isolated package is research/p104-single-pass-qualification/. Its complete official 13-file PR5831 export is retained unchanged at ancestor b41545693de902b2b3b1ffd0218090de0cbcc062; pending predecessor Git objects are not runtime dependencies. The branch carries Ali Romman's same exact test-only PR #5761 dependency, source ba8a9e640e1983f3c29993fefa3f46e8ed03313a, preserving authorship/cherry-pick credit. No other shared-code change is introduced.

Validation

Ubuntu Linux x86_64, CPython 3.12.3. A valid five-event COPY/SCALE/CLEAR fixture uses both real frozen backends, complete fresh output comparisons and 27 input/lifetime/exception/alias controls, normally and under -O, before full construction. The candidate's benchmark includes these controls. Four inherited pre-import custody corruptions remain a separate small prequalification stage inside the registered command.

The first complete pair passes the required 15% wall/CPU gate before four further observations. All six complete scientific receipts match, including scale bits 55 and unchanged prefix/public/owned guards 26/23/22. Baseline wall range is 60.716–63.059 seconds; candidate 38.937–39.017. Kernel peak RSS maxima rise from 101.48 to 115.24 MiB; sampled process-group peaks are reported separately. The host is unreserved, OS caches unflushed and there are three samples per arm: no universal factor or statistical confidence is claimed.

python3 -B scripts/research_package.py --package research/p104-single-pass-qualification

This performs custody controls, new finite controls and fresh full qualification. Complete registered normal and optimized commands pass in 39.240 / 39.179 seconds. The read-only, Git-free implementation export passes under -I -O -B, empty PATH and an explicit authenticated import root in 39.244 seconds; all 91 exported files keep their bytes/modes. The inherited private nested scratch remains writable; this adds no new nested-scratch enforcement claim.

All 15 local guarded commands pass, including both profiles, four finite prototype/hardening commands, six balanced measurements and three full normal/optimized/export checks. They total 533.409 seconds, below the shared 1,500-second cap. The original precision command remains 180 seconds; helper profiling is capped at 300 seconds, sampled group RSS at 1 GiB. Full runs are serialized, with no automatic retries. A draft metadata outcome enum was rejected before registered execution and corrected; its message is retained. No scientific input, expected result, source or resource limit was repaired.

Publication 66fb6c6b8683b37967e9d7f313b3df99a49d2406: all 97 committed package files/bytes/modes, immediate own-input ancestry and official export/input coverage pass. Implementation export 9b8cffa57c09eefba39dc6a4efdc34a15dfc1ebd contains the qualified runtime; later changes are evidence/reporting/notices only. All 11 runtime/input files are unchanged. Full-history preflight is unavailable in this shallow checkout. Exact-head hosted CI and independent scientific acceptance remain separate and will be recorded in the canonical handoff.

Raw scientific stdout is unfiltered; only the volatile timing field is excluded from scientific-receipt equality. Public resource receipts omit local script paths. PROFILE.md, COVERAGE.md, PROOF.md, BENCHMARK.json, VALIDATION.json and evidence/ retain the rejected nomination, gate history, measurements, full outputs and limitations.

Credit

Ali Romman supplied the integer backend, scale proof, explicit adaptation, original qualification and macOS measurements. Preserve PR #5826, PR #5819, Adam Gall, ywz65, PR #2878, Mason Hale/PR #5811 and all retained P104/native/compiler/precision notices. roger-sato/Arham Tahir supplied the historical-source-role test lineage. Guard/custody mechanics follow PR #5816, PR #5818, PR #5832 and PR #5839. Original archives/notices are unchanged. New profiling, one-use observation, finite controls and Linux measurements: Chris R / Clear1nsight, with substantial OpenAI Codex assistance; no subagents or independent human/kernel acceptance. Apache-2.0.

Open next questions

Review the complete-output and observation-lifetime argument before adopting the opt-in command. Another Linux host can repeat the bounded six-process comparison. A source-owned explicit output interface could simplify the scoped wrapper, after coordination and fresh source binding. A changed word/backend requires new coverage and finite feasibility; this package gives no automatic transfer. The engineering scope ends at submission/handoff. Kappa research, Work 1/2, feeds, workers and automations remain paused.

[Clear1nsight]

HANDOFF — complete single-pass precision qualification; local qualification and exact-head hosted CI complete.

1. Research discussion / PR / artifact: PR #5848, research/p104-single-pass-qualification/, following PR #5831 and PR #5839. Bounded claim; profile/finite/first-pair update. Coverage, proof, source archive, complete outputs and every measurement are committed.
2. Exact commit: publication 66fb6c6b8683b37967e9d7f313b3df99a49d2406; implementation export 9b8cffa57c09eefba39dc6a4efdc34a15dfc1ebd; own immutable input ancestor b41545693de902b2b3b1ffd0218090de0cbcc062; source PR #5831 7c5eb98e7439d778141d63e801e53d4cd7e1d587; nested PR #5826 58f93cf41751920f3d6766a39caccc2974ce18b0; refreshed canonical main de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8.
3. What is established: all six fresh complete-qualification measurements in B,C,C,B,B,C order pass. Wall median 61.007 → 39.015 s (36.05% less); CPU 60.950 → 38.880 s (36.21% less). Both complete independent arithmetic contracts execute all 25,984 events/799 coordinates; full source results, ideal metadata/weights, actual traces and every signed witness agree. Every original transpose/literal check, six semantic rejections, original 120-program/4,240-event differential and three integer rejections remain. New first-pass observations are one-use, scoped, deep-copied and input-bound, with no persisted or cross-backend answer cache. Before construction, finite feasibility passes with both real backends and 27 new lifetime/input/exception/alias controls normally and under -O. Four inherited source-custody controls remain. Complete registered normal/optimized commands pass in 39.240/39.179 s. Read-only Git-free -I -O -B export passes in 39.244 s; all 91 exported bytes/modes unchanged. All 15 guarded local commands pass in 533.409 s under the shared 1,500 s cap, original 180 s precision limit and 1 GiB sampled group RSS. Full runs are serial; no automatic retries. All 97 final committed package files/bytes/modes, immediate own-input ancestry, official export/input integrity and all 11 qualified runtime/input files pass. Shared helper setup is only 0.287 of 51.317 s: the fixture-only optimization is rejected and no shared test/wait/skip changes are made. The full ten-test nominated group passes after complete fixture/history authentication. Hosted run 40455483590 passes at this exact head, tested at merge ref b6d2d82dd82ff78782f09cf03e7c1357cf416baf against unchanged main. Its plan passes all 513 tests in 152.788 s with 22 platform skips, selecting only this package. The fresh registered replay passes all four custody controls, 27 finite observation controls and full actual scientific qualification; its entire receipt equals the local result, with inner qualification 38.800884 s and roughly 39.6 s complete command. Committed inputs remain unchanged.
4. What remains uncertain: three observations per arm on an unreserved Linux host with unflushed OS caches are local measurements, not universal speed/statistical evidence. Kernel peak RSS maxima rise 101.48→115.24 MiB; sampled group peaks are separate. The B/C stage includes all qualification and 27 new controls; four inherited custody controls are a separate small stage included in registered/export times. The module-factory wrapper is source-specific, restored on failure/success and documented; deterministic duplicate elimination is not a second hardware-fault oracle. Nested private source scratch retains its original writable semantics. Full-history preflight is unavailable in this shallow checkout, distinct from immediate declared-input ancestry checks. A draft metadata enum rejection and its non-scientific correction are retained. Public resource receipts omit local script paths; raw scientific stdout is unfiltered, with only its volatile time excluded from scientific-receipt equality. Assisted review and hosted tooling CI are not independent human/kernel or supplier acceptance. Five framework assumptions, physical/reset/owned-return/tape/compiler/outer obligations and selected kappa remain unchanged.
5. Smallest useful next task: review the complete-output/lifetime proof before adopting the opt-in command, or repeat the bounded six-process benchmark on another Linux host. A source-owned explicit first-pass output interface might simplify the scoped wrapper, after coordination. A changed word/backend needs new binding, coverage and finite feasibility. This releases the bounded engineering scope and starts no successor; kappa research, Work 1/2, feeds, workers and automations remain paused.
6. Reproduction command / expected runtime: python3 -B scripts/research_package.py --package research/p104-single-pass-qualification runs four source-custody controls,27 new finite controls and fresh full actual qualification with both complete arithmetic implementations and original finite differential, about 40 seconds locally. README documents benchmark.py pair then remaining, preserving the first-pair 15% wall/CPU gate, six-process order, all output/measurement receipts, fixed source/caps, serial lock and no automatic retries. The registered source timeout remains 180 seconds.
7. Predecessor credit: Ali Romman supplied the integer backend, scale proof, explicit adaptation and original qualification. Preserve PR #5826, PR #5819, Adam Gall, ywz 65, PR #2878, Mason Hale/PR #5811, all retained P104/native/compiler/precision notices, roger-sato/Arham Tahir's historical-source-role lineage and exact test-only PR #5761 authorship/cherry-pick credit. Guard/custody mechanics follow PR #5816, PR #5818, PR #5832 and PR #5839. Full archives/notices remain unchanged. New engineering: Chris R / Clear1nsight with substantial OpenAI Codex assistance; no subagents, independent human or proof-kernel acceptance. Apache-2.0.

[Clear1nsight]

CLAIM — bounded verification-engineering follow-up to PR #5848, frozen head 66fb6c6b8683b37967e9d7f313b3df99a49d2406. Fresh full qualification profile passes at 38.314 s; its Fraction signed-error pass takes 20.053 s. Scope: finite feasibility, then at most one source-bound exact Fraction zero/unit fast-path candidate in that routine. This includes immutable constants, coefficients ±1 and magnitude changes 0/±1; all nontrivial arithmetic stays exact Fraction, with the independent integer backend unchanged. Preserve every 25,984 event/799 coordinate result/witness/hash, original 120-program/4,240-event differential, six semantic/three integer rejections, four custody and 27 observation controls. Full actual original/derived trace equality, normal/optimized/sealed checks and >=15% complete qualification wall AND CPU savings are mandatory before a performance PR. Serial fresh processes, first pair before remaining B,C,C,B,B,C, original 180 s command cap, 1 GiB sampled group RSS, 600 s initial stage / 1,500 s shared local cap, no automatic retries. Stop on source/coverage/receipt failure, insufficient complete-stage gain or active overlap. One wrapper-census attempt stopped before target entry; its receipt and explicit 98-member correction are retained and charged. Canonical main remains de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8. Targeted PR/source checks found no matching successor or external PR5848 feedback; this is bounded coverage, not exhaustive/unpublished assurance. Expiry 27 October 2026 at 18:30 AEDT. Preserve Ali Romman and retained PR5831/5826/5819/native/compiler/precision and verification-tooling credit. Chris R / Clear1nsight with substantial OpenAI Codex assistance, no subagents. Kappa research, Work 1/2, feeds and automations remain paused.

[Clear1nsight]

NEGATIVE FINDING / DONE / RELEASED — the exact Fraction zero/unit prototype passes correctness but does not justify constructing a faster complete qualifier. No performance PR or kappa claim.

1. Research discussion / PR / artifact: this follows PR #5848 and the bounded claim. The public feasibility artifact contains the exact prototype, proof, portable finite replay, full actual-audit script, raw evidence, source pins and preserved rejected versions. It is on the fork; no new upstream PR is opened because the required performance gate did not pass.
2. Exact commit: artifact 71f22c77b393e5cc14925c771894dc1a92c8443c; own immutable input ancestor e5614e94d93945963268e76fbd37f7dec58fde78; frozen PR5848 66fb6c6b8683b37967e9d7f313b3df99a49d2406, containing the complete official 97-file export and unchanged PR5831/5826 archives. Canonical main freshly fetched before publication remains de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8. Targeted PR/source-comment searches found no matching successor or external PR5848 feedback; this does not cover unpublished work or establish exhaustive novelty.
3. What is established: fresh complete baseline qualification profiles at 38.314 s, including 20.053 s in Fraction error propagation and only 0.111 s copying outputs. The prototype uses immutable Fraction constants and exact identities for coefficients ±1 and magnitude changes 0/±1; all other arithmetic remains Fraction, and the independent integer backend is unchanged. Before any complete candidate, all 241 finite programs / 10,164 events compare complete values, signed witnesses, numeric types and unchanged inputs; 16 source/code/type/mutation/exception/restoration controls pass normally and under -O. One fresh full actual audit compares 25,984 events / 799 coordinates, all trace/witness values and types, and unchanged inputs. Its complete output digest is 954c7448ff74d78753828e9623ad44c46da9e5ce459420b6e50a5d47294250b9. Within that one process, the original trace takes 20.180 s, derived 19.943 s: a difference of 0.237 s, about 0.62% of the profiled complete qualifier. A 15% whole-stage target needs roughly 5.75 s. The scoped feasibility decision is STOP before complete candidate construction. No full B/C wall/CPU benchmark or performance claim follows. The existing PR5848 qualifier remains the validated baseline. All 61 committed artifact files/bytes/modes, immediate own-input ancestry and official export/input integrity pass; fresh packaged finite replay passes normally and optimized in about 0.9 s.
4. What remains uncertain: this is one ordered within-process trace comparison on an unreserved Linux host, not a fresh-process benchmark, median, CPU saving, statistical effect or universal obstruction to Fraction optimization. Branch/primitive costs were not separately profiled, so the cause of the negligible difference remains open. Complete candidate normal/O/sealed/hosted validation is deliberately not started after the negative gate. No hosted-CI result is claimed for this fork-only artifact. Full-history preflight is unavailable in the shallow checkout. Two attempts stopped before arithmetic comparison: a stale export-member uniqueness count and a code-binding comparison that included nested source locations. Both rejected sources, exact hashes, receipts, diagnoses and separately labelled corrections are retained; the corrections keep all source hashes and recursively bind executable structure. The added nested-code corruption control rejects. Across all nine local guarded attempts, seven pass and those two pre-arithmetic rejections remain visible; total 87.116 s, below the 600-second initial and 1,500-second shared caps, with the original 180-second command limit, 1 GiB sampled group RSS and no automatic retries. Public stderr redacts only the workspace prefix with original/public hashes retained. No failed mathematical differential was repaired. Assisted checks are not independent human/kernel acceptance, supplier admission or a kappa improvement.
5. Smallest useful next task: under a separately authorized scope, profile the remaining Fraction arithmetic primitives before nominating a broader exact algorithm. Repeating zero/unit tweaks or dropping an independent oracle is not justified by this evidence. Any successor needs a new overlap/source/coverage gate and finite feasibility before construction. This claim is released; no successor is dispatched. Kappa research, Work 1/2, feeds, workers and automations remain paused.
6. Reproduction command / expected runtime: from the artifact branch, python3 -B scripts/research_package.py --package research/fraction-zero-unit-feasibility runs a fresh complete finite differential and all 16 controls in about one second. Its manifest explicitly covers finite feasibility only; it does not freshly rerun or replace full qualification. actual_audit.py is separately opt-in, about 43 seconds here, requiring a fresh serial budget and the same source/resource caps. The recorded actual audit should not be rerun merely to acknowledge this handoff.
7. Predecessor credit: Ali Romman supplied the original Fraction/error arithmetic and integer backend. Preserve PR #5831, PR #5826, PR #5819, Adam Gall, ywz65, PR #2878, Mason Hale/PR #5811 and all retained P104/native/compiler/precision notices. PR5848/PR #5839 supply the complete qualifier and Linux baseline; guard/custody lineage follows PR #5816, PR #5818, PR #5832 and exact test-only PR #5761 authorship/cherry-pick credit. Full original archives/notices remain unchanged. New feasibility work: Chris R / Clear1nsight with substantial OpenAI Codex assistance; no subagents or independent human/kernel acceptance. Apache-2.0.

[Clear1nsight]

CLAIM — bounded verification follow-up to PR #5848, exact source 66fb6c6b8683b37967e9d7f313b3df99a49d2406. Following the released zero/unit negative finding, one fresh full qualification with cProfile only inside its actual Fraction trace passes. It records about 26 million Fraction-result allocations, 10.4 million additions, 7.9 million multiplications and 2.58 million coefficient updates. Instrumented costs are not performance evidence. Nomination: at most one source-bound per-trace bounded FIFO table of pure exact Fraction coefficient transitions and weighted increments, complete canonical keys, original Fraction arithmetic on misses; no integer-engine, word, oracle-count, guard or frozen-source edits and no cross-run/backend answer reuse. First test finite values/types, all operation/sign/cancellation cases, invalidation, capacity/eviction, mutation, exceptions and restoration, then one full actual trace audit. Construct a complete qualifier only with credible 15% whole-stage headroom. A performance PR requires >=15% complete wall AND CPU savings in a fresh first pair before all six B,C,C,B,B,C medians, plus normal/optimized/sealed and exact-head CI gates. Preserve all 25,984 events/799 coordinates, full witnesses/weights/metadata,120-program/4,240-event differential, original 6 semantic/3 integer/4 custody/27 observation controls. Serial, no automatic retries; 180 s/source command, 1 GiB sampled group RSS, 600 s initial and 1,500 s shared local cap. Stop on source/coverage/control/receipt/resource failure, insufficient gain or active overlap. Canonical main freshly checked remains de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8. Targeted PR/source-comment/cache searches found no matching successor or external feedback; bounded coverage is not exhaustive/unpublished assurance. Expiry 27 October 2026 at 19:00 AEDT. Retain Ali Romman and all original precision/native/compiler/tooling credit, including PR5831/5826/5819/5839/5816/5818/5832/5761. Chris R / Clear1nsight with substantial OpenAI Codex assistance; no subagents. Kappa research, Work 1/2, feeds and automations remain paused.

[Clear1nsight]

FINDING / DONE / RELEASED — Fraction primitive profiling identifies a useful exact local table, but the complete-cost screen does not clear the requested 15% gate. No performance PR or kappa claim.

1. Research discussion / PR / artifact: follow-up to PR #5848, the bounded primitive-profile claim and the released zero/unit finding. Public code/proof/finite replay/profile/evidence is on the fork. No new upstream PR is opened after the negative complete-cost screen.
2. Exact commit: artifact df9dc4c89b31494f1390d16b9ca2a15e5946e853; own immutable input ancestor 0b2de67914ff659b0e1fca4ff4f275228aa246d6; frozen PR5848 66fb6c6b8683b37967e9d7f313b3df99a49d2406, complete 97-file export including unchanged PR5831/5826 archives/notices. Fresh canonical main at intake and publication is de5ec2e7a6074068452f45b1a191b5c72832df4a, priority 8. Targeted source/comment/Fraction/cache searches found no matching successor or external PR5848 feedback; this is bounded coverage, not exhaustive/unpublished assurance.
3. What is established: one full unchanged qualification passes with cProfile only around its actual Fraction trace. That trace records 26,030,845 Fraction-result constructor calls, 10,396,952 additions, 7,892,384 multiplications and 2,579,750 coefficient updates. The profiled trace/whole times are 65.737/86.794 s and include instrumentation overhead; they are not benchmarks. The candidate local table returns the exact new coefficient and three increments, keyed by all five canonical Fraction operand values. Misses use the original Fraction operators; every caller score/row update, rounded site, full witness/hash and independent integer oracle remains. Tables are fresh per trace, with no cross-backend/call answer reuse. Before construction, 241 programs / 10,164 events at four capacities (964 full comparisons) preserve all values/types/inputs; 34 key/collision/capacity/eviction/source/exception/mutation/restoration controls pass normally and optimized. Two full actual audits preserve every trace/witness value/type and input on 25,984 events / 799 coordinates, with the unchanged full output digest 954c7448ff74d78753828e9623ad44c46da9e5ce459420b6e50a5d47294250b9. The 2,048-entry table gives 19.996→16.578 s with 75.0% hits/643,166 evictions. After an explicit preconstruction reassessment, one final 16,384-entry OrderedDict FIFO test gives 20.251→14.600 s, 87.3% hits/310,400 evictions. Both source versions and all observations remain. The observed 5.652 s trace difference is only 14.49% of the existing 39.015 s complete median before new checks. Adding the observed 1.621 s new finite stage gives a modelled 34.984 s / 10.33% saving, short of the required 33.163 s / 15%. STOP before complete qualifier construction. No complete B/C wall/CPU pair, six-process expansion or performance claim follows. All ten guarded commands pass in 179.502 s, under the 600 s initial/1,500 s shared caps, original 180 s source limit, 1 GiB sampled group RSS and zero automatic retries. All source bytes/modes remain unchanged. All 66 committed artifact files/bytes/modes, immediate input ancestry and official export/input integrity pass; fresh packaged finite replay passes normally and optimized.
4. What remains uncertain: the actual trace times are single ordered within-process comparisons on an unreserved Linux host, not fresh-process medians, CPU gains, statistical estimates or a causal separation of capacity/container effects. The complete-cost screen combines observations; it is not a measured full wall/CPU result or universal obstruction. Complete candidate normal/O/sealed/hosted gates are deliberately not started after the closed screen. No hosted-CI result is claimed for the fork-only artifact. Full-history preflight is unavailable in the shallow checkout. Table entry bounds are not universal byte bounds for arbitrary-size inputs; sampled RSS, kernel peaks and exact source scope remain explicit. The final finite harness derives four unchanged functions once, then performs the same 964 full comparisons and explicitly checks fresh table lifetimes; no numerical case was dropped. Independent human/kernel acceptance, physical admission and kappa are not established.
5. Smallest useful next task: separately investigate whether a source-bound well-formed Fraction input contract and internal type-preservation proof allow moving the new table's repeated type validation out of the hot loop while preserving malformed-input fallback and all controls. The implementation performs five operand-type checks per coefficient update, about 12.9 million checks; its actual cost and any gain from relocation are unmeasured. A successor needs new overlap/source/coverage and finite gates before construction, and must pay the complete qualification bill. No further capacity/algorithm trial is dispatched here. Claim released; kappa research, Work 1/2, feeds, workers and automations remain paused.
6. Reproduction command / expected runtime: on the artifact branch, python3 -B scripts/research_package.py --package research/fraction-contribution-table-feasibility runs fresh finite coverage at all four capacities and all 34 controls, roughly 1.6–2.9 s on this host. The manifest explicitly states finite-only coverage. It does not freshly rerun full actual qualification, certify a performance claim or replace the validated PR5848 qualifier. actual_audit.py is a separately opt-in original/16k full actual trace comparison, about 38 s here, requiring a new serial budget with the same 180 s/1 GiB/no-retry contract. Historical profile/audits should not be replayed just to acknowledge this handoff.
7. Predecessor credit: Ali Romman supplied the original Fraction/error arithmetic and integer backend. Preserve PR #5831, PR #5826, PR #5819, Adam Gall, ywz65, PR #2878, Mason Hale/PR #5811 and every retained native/compiler/precision notice. PR5848/PR #5839 supply the qualifier/Linux baseline; PR #5816, PR #5818, PR #5832 and exact test-only PR #5761 retain tooling/authorship/cherry-pick credit. The prior negative artifact 71f22c77b393e5cc14925c771894dc1a92c8443c supplies source-binding and fixture lineage. New work by Chris R / Clear1nsight with substantial OpenAI Codex assistance; no subagents or independent human/kernel acceptance. Memoization is standard. Apache-2.0.