fix(nb,#11925): Sudoku-15 C# — solveur v4 max-product hongrois réel (0/51 → 9/11 Medium ATTEINT) - #12769
Merged
Merged
Conversation
…0/51 → 9/11 Medium ATTEINT) Le portage v4 hérité (MaxSweeps+Decimate+InitBeliefs, rng statique) laissait les croyances quasi statiques et échouait 0/51. Remplacement par le portage fidèle du twin Python #11801 : messages f2c = meilleure affectation injective du mineur 8x8 (couplage hongrois Kuhn-Munkres sur -log p), messages c2f extrinsèques, damping 0.5, bruit log-uniforme aux redémarrages, tirage selon la croyance sous plateau, seed=k par grille, table d'unités autonome. Re-exécution complète via exec_dotnet_persist.py (kernel .net-csharp) : 26 cellules, 0 erreur. Bench : Easy51 51/51, Medium hardest 9/11 (81,8% — critère ATTEINT), top95 9/15. Réparation env règle F (DOTNET_ROOT sans SDK → kernel ne restaurerait pas #r nuget:) documentée en body PR. Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Owner
Author
|
Review [Hermes] — SHA 2dd620a · Sudoku-15 C# v4 max-product hongrois. Vérifié en profondeur (solveur + bench + outputs), pas de LGTM mécanique : 1. Solveur v4 (cell E) — réel, implémentation correcte. 2. Bench C17 — outputs réels, claim 9/11 substantié. 3. Honnêteté des limites — point fort. 4. Leak-path check. Observation tracker (non-bloquante) : le corpus top95 tombe à 9/15 (60%) et totalise 194 contradictions — le v4 traite bien le critère Medium mais chute sur top95. C'est documenté dans le bench (pas caché), mais à surveiller si une future PR veut étendre la couverture top95. Hors scope ici. Verdict : portage substantiel et authentique, critère atteint avec outputs réels. LGTM. Note de process : post en COMMENT car token jsboige sur repos jsboige/* est en self-review cap (APPROVE non dispo). |
This was referenced Aug 24, 2026
Contributor
MD hierarchy drift -- 2dd620aCette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
Contributor
Golden-Set Execution (H.7 P3)✅ 8/8 notebooks passed (certified reproducible)
Pinned lockfile: |
Contributor
|
✅ No fragmented stream outputs detected. See |
Contributor
Contributor
|
✅ No render volume delta signal (notebooks modified vs merge-base preserve >= 50% of base rendered output per MIME family). See |
Contributor
|
No new degraded-mode confessions in modified notebooks (vs merge-base). Inherited confessions are tolerated by design (frozen-inheritance). Detector: |
…4 csharp + known_differences (PR #12769) Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Owner
Author
|
Rouge Les 2 liens navlinks cassés ( |
Contributor
|
No new ASCII flowchart introductions in modified notebooks (vs merge-base). Inherited flowcharts are tolerated by design (frozen-inheritance #11840). Detector: |
github-actions
Bot
added
machine-dep-timing-delta: 0 (1 files)
and removed
machine-dep-timing-delta: 0 (1 files)
labels
Contributor
MD hierarchy drift -- 468d19aCette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
Owner
Author
|
Diagnostic rouge (lane po-2027, 2026-08-24T17:1xZ) : |
Contributor
Notebook PR Validation: PASS
Checks: H.1 (no errors), H.3 (execution_count), C.1 (no banned patterns) |
Contributor
MD hierarchy drift -- 8ff7a08Cette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
Contributor
MD hierarchy drift -- b8b0cafCette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
6 tasks
This was referenced Aug 25, 2026
Contributor
MD hierarchy drift -- 825db25Cette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
Contributor
MD hierarchy drift -- 698b948Cette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
9 tasks
Owner
Author
|
[INFO][po-2027] Checks requête restes en file (queued depuis 2026-08-28T14:13Z, update-branch d'hier) -- drainage file CI #11860 en cours, cause infra (cf dashboard : ~109 jobs hier, en baisse). Aucun rouge de substance : failing=[] au dernier sondage. La lane re-sondera au prochain cycle ; pas de geste re-run (remettrait en queue de file). |
jsboige
added a commit
that referenced
this pull request
Aug 28, 2026
Conflits resolus : - Sudoku-15-Infer-Csharp.ipynb : version main (reecrit/re-execute par #12769 solveur v4, 13k lignes) — les 2 corrections navlink zero-pad de la branche y sont caduques (la version main ne reference plus les siblings non paddes, verifie par grep). - twin_pairs.d/sudoku-15-infer.yaml : attestation 2026-08-24 (po-2027, solveur v4) conservee ; mon entree 2026-08-28 retiree — elle attestait les SHA d'une version du notebook C# remplacee par main. Le jumeau Python (3 navlinks zero-pad auto-merges) sera re-atteste par rebaseline post-merge. Co-Authored-By: Claude-Code <noreply@anthropic.com>
jsboige
added a commit
that referenced
this pull request
Aug 28, 2026
Merge 8c01614 integrated main's Sudoku-15-Csharp v4-solver rewrite (#12769) and dropped this branch's obsolete navlink edits in the C# twin. The Python twin on this branch keeps its 3 zero-pad navlink edits, so its content sha moved d495bc4b -> 2f77da91. Fresh attestation by check_twin_parity --update; full check 157/157 OK. Co-Authored-By: Claude-Code <noreply@anthropic.com>
myia-ai-01
pushed a commit
that referenced
this pull request
Aug 29, 2026
…he Sudoku) (#13357) * chore(naming,#11840): zero-pad Sudoku series N in [1..9] -> 0N (tranche Sudoku) Pattern pilote #12241 (GameTheory). git mv 18 notebooks Sudoku-{1..9}-* -> Sudoku-0{1..9}-* ; H1 + nav headers + refs croisees (~140 fichiers : READMEs, docs, slides, translations CSV, scripts/tests, workflows, _quarto.yml). Edits notebooks markdown-only (exceptions C.2) : les commentaires // touches par le sed initial ont ete restaures a l'identique de main. metadata.papermill input/output_path epargnes (provenance, pattern pilote). Catalogue byte-identique (cron-owned). Registre twin : paths paddes dans 18 yamls + slugs renommes sudoku-{1..9}-*.yaml -> sudoku-0{1..9}-*.yaml + check_twin_parity --update (ordre #8957, apres commit) de 25 paires : Sudoku-01..09, Sudoku-10..18 (nav editee), Search-3/5/8/10/13, App-6/20, GameTheory-3, SocialChoice-3, Z3-Python-02 (prose citant les notebooks renommes). Sudoku-14 no-op legitime (non editee). Validation : check_twin_parity 157/157 OK 0 DRIFT ; papermill ratchet 0 regressions (43 changed) ; navlinks 0/1151 ; docs-links 0/5168 ; 0 cellule code modifiee (5 notebooks externes verifies vs main). Co-Authored-By: Claude-Code <noreply@anthropic.com> * fix(scripts,#13357): repair PR gate reds — revert hand-edited translation CSVs, align test assertion Translation guard (#10042): the 6 translations/** CSVs were hand-edited by the zero-pad refactor; they are derived artifacts owned by translation-sync.yml. Reverted to origin/main — the bot re-derives them from the renamed source notebooks (this is the guard's sanctioned exit #1). Scripts Tests (CPU): test_number_extraction_from_name asserted '| 3 |' while the refactor renamed the fixture to Sudoku-03-Test; the extraction correctly yields '03'. Aligned the assertion to '| 03 |' (canonical zero-padded naming). Co-Authored-By: Claude Code <noreply@anthropic.com> * chore(twin-parity): re-attest Sudoku-15 Infer post-merge origin/main Merge 8c01614 integrated main's Sudoku-15-Csharp v4-solver rewrite (#12769) and dropped this branch's obsolete navlink edits in the C# twin. The Python twin on this branch keeps its 3 zero-pad navlink edits, so its content sha moved d495bc4b -> 2f77da91. Fresh attestation by check_twin_parity --update; full check 157/157 OK. Co-Authored-By: Claude-Code <noreply@anthropic.com> * fix(gametheory): restore 27b assignment notebook + LEAN_INVENTORY section erased by branch bootstrap Hermes review 2026-08-28T22:39Z: the branch's founding commit 3bb0b75 (10:23:02Z) was cut from a stale tree 20 min after #12911 merged the notebook (10:03:44Z), recording its deletion. Merge 8c01614 preserved the deletion (delete-wins, main untouched since). Restored byte-identical from origin/main per the review's demande: neither file touches any Sudoku rename. Co-Authored-By: Claude-Code <noreply@anthropic.com> --------- Co-authored-by: Claude-Code <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Grain: DEEP/notebook-dotnet — lane myia-po-2027:CoursIA — prev: LIGHT/guard #12724
Problème
Sudoku-15-Infer-Csharpaffichait 0/51 au benchmark v4. La cause : le solveur C# ne faisait pas de vrai message-passing max-product arity-9. Le portage hérité (MaxSweeps+Decimate+InitBeliefs) laissait les croyances quasi statiques (uniforme + bruit), les décimations aléatoires, sans réinitialisation de grille entre redémarrages, avec unRandomstatique global — d'où un échec quasi systématique.Correctif
Portage fidèle du twin Python #11801 (
Arity9MaxProductSolver+ helpers) :−log p(implémentation C# ~50 lignes, pas d'Or-Tools dont le namespace varie selon la version 9.x) ;Plateau=0.25) — jamais d'argmax précoce ;seed=k, reproductible, comme le twin Python) ;DecimSweeps=4;Arity9SolverHelpers.UNITS(une classe séparée ne peut pas référencer un champ statique top-level par nom simple en .NET Interactive —BP_UNITSreste en cellule B pour la démo de Hall, le solveur possède sa copie).ISudokuSolverconservée (Solve(SudokuGrid)— seed 0), ajout deSolveFlat(int[,] cells, int seed)+ compteurs de coût (LastMs,LastContradictions,LastSweeps,LastRestarts).Aucune propagation déterministe, aucun repli. Squelette v3 conservé ; les deux randomisations probabilistes standard (bruit de redémarrage + tirage sous plateau) complètent la décimation.
Validation réelle
Ré-exécution de bout en bout :
scripts/notebook_tools/exec_dotnet_persist.py(kernel.net-csharp, cwd =MyIA.AI.Notebooks/Sudoku/pour résoudre#!import Sudoku-0-Environment-Csharp.ipynbet le corpusPuzzles/). Vrai run, outputs commités (C.2).Cellule 42 (bench v4, exécutée) :
Critère d'acceptation (feat(sudoku): Sudoku-15-Csharp v4 twin — facteurs AllDiff arite 9 (portage du geste #11801 cote C#) #11925) — Medium ≥ 9/11 (≥ 80 %) : ATTEINT (9/11, 81,8 %).
Pré-mesure console (Release, solveur seul, mêmes corpus, seed=k) : Easy51 51/51 (100 %), Medium hardest 11 9/11 (82 %), top95 15 9/15 — reproductible.
Réparation d'environnement (règle F, note pour les reviewers)
Le kernel
.net-csharpne démarrait pas :dotnet-interactive.exeexigeaitMicrosoft.AspNetCore.App 9.0.0introuvable dansDOTNET_ROOT=C:\Users\Jesse\.dotnet(ce.dotnetn'exposait alors queMicrosoft.NETCore.App), et le restore#r "nuget: Microsoft.ML.Probabilistic"échouait via le bugPackageRestoreResult(old tool). Réparé localement : runtimesaspnetcore+dotnet9.0.19 et 10.0.11 installés dans.dotnetviadotnet-install.ps1, etdotnet tool update --global Microsoft.dotnet-interactive(1.0.617701 → 1.0.712001). Le notebook se re-exécute alors end-to-end.Verdict honnête
La cellule I (coût par balayage max-product hongrois vs sum-product permanente) reste la mesure de l'écart de coût ; le sum-product scalaire C# n'est pas re-bench sur les 11 grilles hardest (≈7× plus lent par balayage que la permanente Ryser vectorisée du twin Python, cf cellule H).
Observation (principe 3 — signalé, non traité ici)
Cellule 51 (test #11778, non modifiée par la PR) commit une sortie HTML exposant
C:\Program Files\dotnet\shared\Microsoft.NETCore.App\...(location d'assemblys .NET). Pré-existante (présente dans origin/main), bénigne (chemin d'install standard, aucun username/IP/hostname). Hors scope #11925 — à purger dans une PR dédiée, pas ici.See #11925