Add categorized experiment log: 100 ideas, 300 trials, and recorded failures

[rohanarun]

Latest completion: the coordinated partial-own candidate now passes local physics, explicit owner matching, global controller/invoice and exact arithmetic at conditional kappa0.001388082565814657 (PR473). negative-searches/partial-own-coordinated/ retains the initial infeasible allocation, missing-input launcher failure, full-model timeout without incumbent, restricted-model repair, and successful verification receipts. A failed fixed-pattern allocation does not prove the gauge family infeasible.

Latest addition: negative-searches/partial-own-coordinated/ records fresh local verification of the four-gauge partial-own word and a rejected aggregate allocation. The stricter pattern/call flow is infeasible; the optimistic floating root is not an admitted bound. Solver timeout and downstream missing-input failure are preserved.

Research results and failed attempts were spread across PRs and SSD folders, making it easy to repeat exhausted searches or mistake execution errors for mathematical failures. This PR keeps research/experiment-log/ as the shared categorized record, with evidence, scope, baselines and retry guidance. Numerical improvements remain in #473; this PR claims no new bound.

The initial import preserves all 100 campaign ideas across eight topic files, including 40 ideas with no execution evidence; three separate 100-trial kernel batches; 14 other scoped searches; and 16 execution-failure attempt groups with 31 original artifacts. These counts are the import baseline, not the total of later followups. Search iterations are not full mathematical certifications.

Latest additions:

• Eight donor-expression runs with no improving moves, plus four basis-exchange runs totaling 20,000 attempts. All endpoints, seeds, histories, exact prices and losing alternatives are retained. The successful rank-139 construction is published separately in Conditional kappa 0.001467289440740487: rank-sector factory on 112 ports #473.
• Four further runs totaling 80,000 exchange attempts with temporary entrance demotions: no global improvement. One candidate improved the local search score but lost after exact whole-family pricing. Its word, price, history and retry guidance are preserved.
• The cancellation-first hypothesis: paired source-injection removal did not change helper first frames. Local replay passed, but it supplied no new entrance opportunity.
• All 424 hierarchy alternatives plus the baseline, subsequent R714 local replay and 40 no-gain kernel trials, five PR485 composition variants, and both initial replica-28 manifest failures with their correction.
• Rank-139 validation receipts: fresh full Python 3.14 replay rebuilt both suppliers; Python 3.11 completed with explicitly pinned, completed-complex reuse. These are distinct verification scopes.

Records distinguish mathematical rejection, no improvement, solver/model limitations, execution failures, unverified proposals and completed verification. Floating packing-marginal diagnostics are labeled as discovery aids, not exact dual optimality certificates. Original failures remain alongside corrections; agents should document a changed input, assumption or search neighborhood before repeating a completed experiment.

Validation: archive hashes, all 100 original idea IDs, all 300 original batch trial IDs and local Markdown links pass. Publication is checked against GitHub blob hashes. Historical mathematical claims are not claimed to have been rerun. Original campaign records remain linked to their pinned PR368 archive.

Further complex-side results: PR496 packing at 8/16/32/64 replicas, 106 entrance cases, 112 rational basis-exchange cases, and five retiming screens all found no admitted improvement. Original execution failures and the corrected runs are retained. A tiny fallback-count arithmetic proposal is explicitly unreviewed for complex compiler applicability. Fresh Python 3.14 headline/anchor complex-provider replay and the subsequent six pooled-component stages passed; no full two-supplier PR496 replay is claimed.

Newest followups: 476 coordinated PR496 gauge trials (128 local constructor passes, 348 rejections, 22 profile classes), all-hosted and bank-permitting packing with no gain. Against the higher PR501/502 retimed baseline: 6,630 directed frame groups, a 400-move neutral walk, and six stage-count models also find no improvement. Frozen-map reconstruction matches the pinned PR501 word; its initial Python runtime failure is recorded separately. These are explicitly scoped search screens, not complete new-supplier certifications.

New gauge/kernel category:52 direct entrance rejections;1,253 selective gauge trials (1,121 local constructor passes,132 rejections,65 profile classes);186 kernel/gauge trades (124 passes,62 rejections,79 classes); both homogeneous pricing screens and two16-replica extensions find no gain. A143-class stage-dependent allocation found a small exception, independently admitted in PR473 and subsequently surpassed by PR505. Full ledgers, frozen inputs, failed attempts, exact arithmetic, completed followup receipt and retry guidance are preserved. These counts describe search trials, not independent complete theorem verifications.

Asynchronous followups:10,224 fixed-map substitutions all exceed physical capacity; baseline and143-word fixed-order integer allocation screens give zero gain. The reference plus48 new replica schedules are retained, with four integer-priced nominations. The best candidate's initial per-pattern matching failure and its stricter global-flow repair are preserved, alongside40,320 explicit owners and completed conditional finite-composition receipts. The new numerical improvement is in #473. An initial list-versus-map execution error is archived separately.

Exact fixed-inventory wall:133 distinct timing schedules and the143-word mixing LP find no gain. A new rational dual certificate checks all348 width patterns and146 frozen word profiles using two independent interval engines. At the next complex saving grid point both prove moment>1, and the retained six-step outer assembly cannot reach the next kappa grid point. This excludes only the specified inventory/menu/fallback/assembly model; it is not a universal bound. Three new retimed kernel selections,16084 rank-two entrance subspaces and292 donor/pivot exchanges are also recorded, with no gain. The portable exact certificate has been replayed from the staged archive. New factories such as #507 are outside this obstruction.

Partial-own followup: PR507's new factory lies outside the146-profile obstruction. New allocation, stronger pattern/call flow and literal matching produce a fully checked conditional composition in #473; discovery records and the completed receipt are retained here. A later transfer of #508's coordinated four-gauge idea passes only the local constructor and is explicitly unverified as an additional global gain.

[JorgeCepeda]

Missing spaces in description.