[petalas]

Summary

Supply the finite combined-word precision implication left open by PR2139 and PR2096 question 2. All changes are under research/p104-delayed-precision/.

Completed forward P104 identity word
  exact all-cut input coefficient P = 131072
  exact absolute-defect coefficient A = 1394705
  exact weighted-defect coefficient K = 2134050
  sufficient state-relative guard = 39 bits

The ideal endpoint maximum is only two. All three interior peaks occur after event 33373 in source93, whose input row is -65536*(input86+input93). Input86 is an arbitrary dirty helper. Endpoint-only bounds therefore miss the transient.

Under the stated completed-instruction Hilbert defect contract, u <= 2^(-b-39) suffices for all-cut error at most B*2^-b, for every integer b >= 0. This is a sufficient local feedback estimate, not minimal numerical precision or an actual rounded-child implementation.

PR2139's complete construction is retained unchanged, including its unfavorable fixed-profile cost screen. No supplier, precision-charge reduction or new kappa is claimed.

Evidence

Before this contribution, PR2139 supplies a completed forward identity word but leaves precision open. PR2096's constants cover the different G/2 word and cannot be transferred here. After this contribution, the registered replay checks the complete combined stream and its signed all-cut sensitivity maps.

Observed on CPython 3.12.15, Linux x86_64, standard library only:

python3.12 scripts/research_package.py --package research/p104-delayed-precision
python3.12 -W error -B -m unittest discover -s tests -p 'test_research_*.py'
python3 -B scripts/research_baseline.py

• Every exported stream field regenerates from the unchanged PR2139 producer; its 241 new native recipes, 421 occurrence bindings and phase fault control pass.
• All 844 initial columns, every cut, arbitrary dirty restoration, private resets and the complete profile and rank mass 6782 pass. Cached norms are checked against direct coefficient-map reductions at every update.
• Separate Fraction reverse adjoints agree at the peak cuts and seven endpoints; sign-aligned executions attain the three coefficients.
• Nine behavior groups pass normally and with assertions disabled. Controls reject missing compensation/clear, illegal charts, corrupted receipts/words, changed producer code before import, capacity violations and overwrite attempts. Relocated default replay is read-only and needs no Git/network.
• All 65 repository research-workflow tests and selected source-binding orientation pass. This is not the selected cold replay or a hosted CI pass.

The sole task draw was p104-side-correction, nonce 159427fb87da401855d7b74679378fb5, priority revision 4. TASK.json explicitly changes scope to this precision implication rather than another correction construction. Base and refreshed canonical main are 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2. The final overlap check at 21:31 UTC covered 415 open PRs and 837 Discussion records with complete pagination and no warnings, not all diffs or unpublished work.

Credit Jose Manuel / PR2139, pinned at 1df674ca8e9d9e0fb98ba4df402b648e9b3be0a4, for the entire completed construction. Credit Adam Gall / PR2096, pinned at f22626f334a06f6e842766e4a6647419e6821d2d, for the Hilbert-defect and feedback method on the older channel. The Rohan Arun, Chafik Boukhalfa, Ali Romman, Adam Gall, Josie Elliston, Clifford and selected Hamming/dirty-helper lineage is retained in NOTICE.md and the original predecessor files.

Coordination and handoff: Discussion2121 claim.

Merge danger

Door: Two-way. This is an additive scoped research package with opt-in replay.

Blast radius: Local. No selected record, frozen predecessor, existing verifier or priority catalogue changes.

The native unitary-chart and completed-child defect contracts remain premises. Rounded recursive children, a fresh audit of every retained native recipe, global owners, the dual word, tape/all-size admission and independent human review remain open. The package includes concrete next questions; it does not establish a complete multiplication theorem.

[ywz65]

CLAIM - operand-weighted state-relative feedback for the exact completed P104 word. Account ywz65; substantial OpenAI Codex assistance. Starting current canonical main c3c94c2, priority5; baseline bindings PASS. Exact source2166 a655734 unchanged/no prior comments. Instead of multiplying the global ideal peak P by the global weighted-defect peak K, compute each site ideal operand norm a_j from the unchanged signed input rows, then J=max_(cut,row) sum_j |c_rj| a_j. The candidate theorem is E <= u*(JB+KE), with full source, all-cut and backward-covector verification. Stop if it gives no material uniform-guard improvement or overlaps a completed result. This is a distinct refinement of this exact word, not a new decoder or native rounding implementation. PR2670 is credited for the simpler QQ16 K(P+1) integer-guard audit; PR2654 and PR2549 concern different P104 words. Exact reset/unitary-chart/completed-instruction assumptions remain, including arbitrary dirty inputs and all34334 boundaries. No selected precision bill, unfavorable source cost screen, supplier or kappa changes. Cached767openPRs at18:06UTC and currentmain reviewed; Discussion snapshot stale due rate limit. This narrowed recursive-interface precision obligation replaces the retired duplicate scratch-slice scope; no duplicate contribution is counted.

[ywz65]

DONE / claim released: #2726 at e10b650. Exact operand-weighted J=66,657,328 and retainedK=2,134,050 give E<=uJB/(1-uK), sufficient27uniformguardbits and26forb>=3. Fulloriginalreceipt, all-cut scan,844idealendpoints,7reversecovectors andsemantic/exportcontrols pass. Sourceword/nativeinvoice remain unchanged; unitary/exactreset/completed-output contract retained. No actualroundedchild, minimalphysicalprecision, supplier or newkappa claim. Credit PR2166 and its complete lineage; substantial OpenAI Codex assistance.