Remove P104 private descents through late FULL dirty-query delivery

research/p104-dirty-catalysis/PROOF.md

@@ -0,0 +1,158 @@
+# Late FULL delivery removes the eight private descents
+
+## Contract and changed premise
+
+The object is the exact 4,058-event witness frozen in the main-branch native
+P104 package, followed by the actual PR3882 correction builder at scientific
+head `9d56eeedbed5e3e9bbbc8c189d66f14cf15b6067` (publication
+`d76586303e7036baa74f8deb2155038290174ffd`). Inputs are 104 source rows,
+104 target rows, and 72 independent arbitrary dirty Q/G/D rows. Coefficients
+are dyadic; no query or delivery helper is initialized to zero.
+
+PR3965 establishes that deleting the eight copy intervals with their original
+consumers unchanged is wrong, and that replacing each private descent by a
+same-query descent/return pays another56. Here we change the consumers and
+cleanup jointly. We retain every public native transport and the original
+Q producers, initial compensations, and Q source cleanup. We replace the
+32 early copy deliveries **and** their32 paired D cleanup reads by416 late
+FULL reads. The104 target lifts already present in PR3882 are retimed, once.
+This is a different chronological contract, not a counterexample to3965's
+fixed-consumer result or the earlier fixed-boundary COPY56 results.
+
+## Actual cones and causal identity
+
+For each query j the verifier extracts the literal interval
+`zero C; copy C,Qj; transport C; Dk += C ...; erase C`, its D consumers,
+their later Y reads, and their FULL cleanup `Dk -= Qj`.
+The eight query records in RESULT.json contain every original zero-based
+index, all416 real Y reads, all32 cleanup indices, and all query source
+coefficients. Each query at copy time is exactly
+
+    q_j = h_j + a_j^t x,
+
+with coefficient1 on its own arbitrary dirty input h_j and no other dirty
+input. The actual producer a_j is extracted from the whole word; it is not
+postulated. It stays untouched, apart from native frame changes, until cut
+3641. Each Y response u_ij is the sum of actual post-injection reads of
+that query's D helpers. There are52 nonzero u_ij per query, all ±1/4.
+
+Each D helper's initial dirty contribution to Y already has its compensating
+initial read. Removing both its q_j injection and its final subtraction
+therefore restores its original dirty row while omitting exactly u_ij q_j
+from each Y row. The inherited initial Q/Y compensation remains. Adding
+u_ij q_j directly at FULL restores the exact old Y row, including cancellation
+of h_j; the existing subsequent Q source cleanup restores q_j to h_j.
+No dirty erasure, missing copy owner, or changed source preparation occurs.
+
+This identity is verified on all280 independent input columns by exact
+Fraction arithmetic. It preserves every final row, including the residual
+encoding, rather than only the104 sum outputs. For a bounded pilot, just
+Q0's five D helpers and52 reads give combined mass2569 and exact equivalence;
+the same verified extraction extends to all8 queries.
+
+## Chronology and native physical operators
+
+Delete the eight complete C intervals and32 D/Q cleanup additions. Retain
+all other witness events. Immediately after original event3641, lift every
+Y from its literal final hyperplane to FULL. All104 X and72 Q/G/D are already
+FULL, all Y have finished their terminal transports, and the remaining416
+original events are only Q/X source-cleanup additions at FULL. For each Qj:
+halve twice, perform its52 signed unit Y additions, then double twice.
+Thus the donor is restored before any Q cleanup. Complete the old Q cleanup.
+Finally execute the old PR3882 tail with its104 initial lifts removed.
+
+The physical row invariant is the native package's
+`physical_r = Gamma_(frame_r) * logical_row_r`. The retained public native
+subsequence is checked literally equal to the inherited sequence. The only
+removed positive children are the eight private rank7 descents. The104
+inserted lifts are implemented in LIFT-RECIPES.json, each as actual Gaussian
+monomials around one rank1 positive child, not admitted by a rank count.
+Their old frame is bound to the **exact ordered canonical basis** in the
+certificate for the corresponding last Y transport; equality of spans alone
+would not fix its chart operator. Their target is the same F9 used by Q/X.
+The independent native interpreter checks signed Pauli actions and exact
+phase for all104 complete operators. These determine the full Clifford;
+512 Gaussian columns are also compared for the first lift. A changed global
+phase rejects. No new direct read is allowed outside the common FULL chart.
+
+The remaining historical native operators and endpoint adapters are inherited
+from PR1991. We do not claim a fresh replay of that entire native certificate.
+The raw changed module now has104 source and target FULL endpoints,72 restored
+dirty FULL helpers, and no private allocation or discard. It implements the
+same Gram row map. Its monotone boundary progress and native mass are both
+2416 (=2312+104). This changed raw boundary is not the old partially transformed
+channel boundary. It does not by itself implement I or admit a global supplier.
+
+The saving is also available to the source-preserving PR3894 continuation
+(head `13b397e36d43f2a91b9e6d42b3ca67f0ec50cea7`). Its104 Y lifts can be
+grouped first, since each target's later correction reads only unchanged X
+rows and its lowering touches only that target. Our raw FULL endpoint equals
+the original raw channel followed by those grouped lifts on all280 columns.
+Thus its remaining4,992 target additions and104 target returns compose with
+the identical input interface, giving the same2576→2520 native saving without
+consuming X. This is a compositional transfer using PR3894's stated schedule;
+we do not claim a fresh replay of its frozen correction here. In particular,
+56 saved native mass is not an advantage exclusive to role exchange.
+
+The final104 drops/role-exchange interface of PR3882 retains that predecessor's
+physical/global admission qualifications. This package establishes exact
+normalized endpoint equality and a new native implementation of the retimed
+prefix; it does not fill the unproved all-size or global-owner implications.
+
+## Complete local accounting and conditional caller gate
+
+| Item | Original channel +3882 tail | Changed module +remaining tail |
+|---|---:|---:|
+| Native channel rank mass |2368|2312|
+| Y lifts |104|104|
+| Final drops |104|104|
+| Total positive-child mass |2576|2520|
+| Private zero/copy/erase |8 each|0|
+| Raw channel weighted ADDs |2624|2976|
+| Explicit new half/double |0|16 each|
+
+Original channel profile is
+`{1:1074,2:58,3:86,4:104,5:48,6:24,7:8,8:8}`.
+Changed raw module including lifts is
+`{1:1178,2:58,3:86,4:104,5:48,6:24,8:8}` (mass2416).
+Add104 rank1 final drops to obtain combined mass2520. The correction's
+remaining1248 unit ADDs and104 each half/double are unchanged.
+
+For a reproducible complete scalar tariff, lower each old coefficient p/2^s
+separately as s donor halves, |p| signed unit ADDs, s donor doubles. This is a
+valid deliberately unoptimized implementation, not a scalar lower bound.
+Old channel:3872 unit ADDs and2496 each half/double. New module:4224 unit ADDs
+and2512 each half/double. Including the unchanged tail gives5472 unit ADDs
+and2616 each half/double, versus5120 and2600. There are no unpriced nonunit
+coefficients in the new416 deliveries. All old widths/precision obligations
+remain; rank mass does not erase this finite scalar overhead.
+
+Under precisely PR3954's optimistic two-axis hosting/coverage assumptions,
+there are208 calls and10816 anchors. Corrected local endpoint progress remains
+2312. The old complete excess44096 decreases by208*56=11648 to32448:
+
+    208*(2520-2312) - 10816 = 32448 > 0.
+
+The complete bill is2997696 over denominator2965248. Its zero-exponent moment
+is therefore greater than1; with all widths at most81, the same monotonicity
+argument excludes the retained favorable gate throughout 0<=a<1. No further
+construction round follows from this gate. The raw all-FULL module's changed
+progress2416 requires a new caller/entrance/coverage invoice; substituting
+2416 into the old formula would silently change the contract. We make no
+such global claim and do not change the selected frontier.
+
+## Verification and remaining work
+
+Four semantic controls remove delivery, move delivery after Q cleanup, omit
+one real target, or fail to restore the quartered dirty donor. Each fails
+exact row equality. Separate optimized CLI processes reject Boolean/float
+receipt types, duplicate keys, false savings, changed/missing source bytes,
+and changed frozen tail bytes, after first validating each private fixture.
+All checks use exceptions, not optimization-removable assertions.
+
+Next genuinely different questions are an actual caller for the new all-FULL
+Gram boundary and a complete jointly redesigned correction with owner/coverage
+and all-size accounting. Combining this change with PR3943's earlier tail and
+compressed Q cleanup requires a new chronology proof; the scalar savings from
+that PR are not credited here. No universal copy or native-rank lower bound,
+new source-preserving cost optimum, or improved kappa is claimed.

research/p104-dirty-catalysis/README.md

@@ -0,0 +1,51 @@
+# P104 late FULL dirty-query delivery
+
+A concrete coupled consumer/cleanup rewrite removes all eight private rank7
+COPY descents while preserving the exact PR3882 endpoint encoding and all72
+arbitrary dirty helper returns. The existing104 target lifts move before Q
+source cleanup;416 late FULL reads replace32 private deliveries and32 paired
+D cleanup reads. Native mass falls2576→2520, with352 additional unit ADDs and
+16 each donor half/double. No private register is created or erased.
+
+The raw all-FULL Gram module has mass and monotone progress2416. It changes the
+boundary of the original2312-progress channel. Under the unchanged PR3954
+corrected two-axis contract the complete gate still fails by32448. No supplier,
+all-size admission, frontier improvement, or universal copy lower bound.
+
+See [PROOF.md](PROOF.md), [RESULT.json](RESULT.json), and [VERIFY.json](VERIFY.json).
+Run `python3.12 -B research/p104-dirty-catalysis/verify.py` and `native.py` at
+the same path. The latter independently checks104 actual rank1 lift operators
+including global phase, bound to the inherited canonical physical charts.
+The full historical native channel is inherited, not freshly replayed here.
+
+## Provenance and scope
+
+Canonical main20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority6. Explicitly
+assigned changed-consumer study. The07:38:11UTC complete676-open-PR body cache
+(including drafts), repository overlap helper, and root coordinator's live
+board/claim review found no equivalent late-FULL delivery construction. The
+helper's Discussion cache was absent; this is bounded overlap evidence, not
+a novelty certificate. Final checkpoint is recorded in the PR handoff.
+
+Credit Rohan Arun's PR1972 scalar/frame word; Chafik Boukhalfa's PR1991 native
+certificate; Adam Gall's PR1940 delivery work; compiler lineage icekylinx PR531
+and independent interpreter PR1862; Josie Elliston, Sagar Thakrar, Douglas
+Colkitt and Dominik Scholz's E8/P104 work. Chris R / Clear1nsight's PR3882
+supplies the unchanged geometry and actual correction builder, PR3943 the
+related earlier-cut Q cleanup precedent, PR3954 the conditional two-axis bill,
+and PR3965 the source-bound fixed-consumer audit and sparse evaluator adapted
+here. PR3961/3962 and PR3966 establish distinct boundary, coverage, and gauge
+constraints. This rewrite changes the producer-to-consumer chronology beyond
+3965's fixed-route premise and credits no3943 scalar cleanup savings.
+
+Prepared by Chris R / Clear1nsight with substantial OpenAI Codex assistance.
+Apache-2.0, following the repository and retained source licenses. Unchanged
+pending-PR builder bytes are vendored with exact original blob/head/hash in
+SOURCES.json; no unmerged pin is assumed to exist in a main-only source export.
+
+## Open next questions
+
+Can an actual caller use the changed all-FULL Gram boundary with a complete
+entrance, physical-intertwiner, arbitrary-dirty-owner, coverage and all-size
+invoice? First specify that caller and its full cost gate. The corrected
+PR3954 caller still fails, so this package stops with the local construction.