docs(repo): #415 Phase-2 affinement — Mindmapper gap (187 MB) + refreshed pack

[jsboige]

Summary

Tertiaire of dispatch kp26j6. Refines the existing #415 Phase-2 proposal (docs/repo/415-phase2-proposal.md, PR #628) with a fresh re-measurement on current master 9c19e51a. Read-only — no history mutation, no force-push, no clone. DoD = update proposal doc; go/no-go remains jsboige's call.

Findings

1. GAP in the proposed --path set — 186.7 MB missed

The proposal's git filter-repo --path prefixes are exact-prefix matches. They miss two sibling regenerable build-output trees:

Missed tree	History	Why missed
Generation/Converters/Argumentum.AssetConverter/Mindmapper/Published/	129.2 MB	.../AssetConverter/Mindmapper/Published ≠ prefix .../AssetConverter/Published
Cartes/Generation/Mindmap/Mindmapper/Published/	57.5 MB	Cartes/Generation/Mindmap ≠ prefix Cartes/Generation/Converters

These are regenerable dotnet publish output (same class as the main Published/): Mindmapper project source (.csproj, .sln, MindMapConfig.cs) is committed in history, binaries deleted at HEAD (0 at HEAD), and .REMOVED.git-id sentinel files alongside confirm the team already un-tracked them as build artifacts. They were misclassified into the other zone by the audit's path regex — which is why the proposal's ~1.557 GB under-counted.

Corrected strip target: ~1.8 GB regenerable (+187 MB vs proposal). The --path set in the plan is updated to close the gap.

2. Pack grew, but bloat source is not recurring

size-pack 2.05 → 2.08 GiB, 3 → 4 packs, +3868 objects — but the growth is text/code only (other zone 882.8 → 946.7 MB). Regenerable binary zones (Published/, Downloads, .resources) are byte-for-byte unchanged. No new build artifacts were committed; the strip target is unchanged in nature.

3. Risk #2 currency: open PRs 8 → 3

At proposal time: 8 open (#627/#603/#596/#597/#599/#600/#601/#627-area). Now: 3 (#596, #661, #662). Lower rewrite risk than at proposal time.

Recommendation

Unchanged: DEFER to post-v0.9.0-tag. Rewriting during the freeze (pending tag + open PRs) maximizes risk for ~zero release benefit. This PR sharpens the plan for when jsboige green-lights it post-release.

Reproducibility (code=truth)

python tools/git-weight-audit.py          # pack + zone aggregation
# scratchpad/415_mindmapper_gap.py        # the 186.7 MB gap quantification

Both read-only (git plumbing). No git filter-repo / BFG / force-push run.

Relates #415, #621, #628.

🤖 Worker po-2024 (dispatch kp26j6, tertiaire)

[clusterManager-Myia]

[NanoClaw]

LGTM — docs(repo) #415 Phase-2 affinement — Mindmapper gap (187 MB) + refreshed pack (head cde6bf3b, +42/-3, docs-only docs/repo/415-phase2-proposal.md, read-only — no mutation/force-push/clone). Firsthand au head :

Affinement sain et bien raisonné (re-mesure sur master courant 9c19e51a, était 33b1c0bc) :

• Pack drift qualifié : size-pack 2.05 → 2.08 GiB, 4 packs, +3 868 objets — la croissance est texte/code only (zone other 882.8 → 946.7 MB), les zones binaires régénérables (Published/, Downloads, .resources) byte-pour-byte inchangées → conclusion correcte « la source de bloat n'est pas récurrente » (pas de nouveaux build artifacts commités).
• 🔴 GAP réel identifié (le cœur du PR) : les préfixes git filter-repo --path de la proposition originale sont des matchs exact-prefix → ils ratent 2 arbres siblings Mindmapper/Published :
  • Generation/Converters/Argumentum.AssetConverter/Mindmapper/Published/ = 129.2 MB (.../AssetConverter/Mindmapper/Published ≠ préfixe .../AssetConverter/Published)
  • Cartes/Generation/Mindmap/Mindmapper/Published/ = 57.5 MB (Cartes/Generation/Mindmap ≠ préfixe Cartes/Generation/Converters)
  • Gap total 186.7 MB qui aurait survécu au rewrite. C'est un vrai gotcha git filter-repo (les --path ne matchent pas les sous-chemins siblings plus profonds).
• Blobs missed : Mindmapper osx-x64 (71.6 MB), Mindmapper.exe win-x64 (57.5 MB ×2 arbres) — régénérables (dotnet publish, 0 à HEAD, sentinels .REMOVED.git-id confirment le un-track). Appartiennent bien au strip.
• Correctif proposé cohérent : ajoute les 2 --path manquants + NOTE explicative. Cible strip corrigée ~1.8 GB (était ~1.557 GB ; le +186.7 MB explains la différence). Arithmétique interne cohérente (1.2 GB Published + 187 MB Mindmapper + 205 MB Downloads + 150 MB resources ≈ 1.74 GB).

Refresh factuel : compte PRs ouverts 8 → 3 (#596, #661, #662 — les autres mergés/clos depuis la proposition d'origine), marqué « Lower risk than at proposal time ». (Note : #663/#665 ouverts par la même dispatch entre-temps — le snapshot doc reste honnête pour son timestamp.)

Scope respecté : DoD = update proposal doc, go/no-go reste jsboige's call. 0 mutation d'historique, 0 push. Review = validité du raisonnement (sound) + cohérence interne (✓), pas re-mesure git (pas cloné le gros repo — bon scope pour un doc de proposition).

— NanoClaw (myia-ai-01) — LGTM