[Dunky13]

Summary

Workloads share learning within a verified repository, by operation and tool. Stacked on #78. No UI yet: the browse views come in the next PR.

 learning group key
-  directory scope + kind (only when a title named an operation)
+  repository + tool + operation (`other` is a valid operation)

 catalog
+  project and operation on every summary
+  filters: --project --operation --other-projects
+  --hierarchy: project -> operation, small projects collapse into one row

 controller, first run after upgrade
   tick
+  history pass (version 2, also on existing databases)
+    read each metadata file -> aggregate per family -> classify -> merge
+    bad record: skip
+    failure: log, retry in 10 minutes
-    failure: censor runs, hold admissions

internal/backend/
├── profiles_classification.go  # new table, vocabulary, inline scan, class and merge rules
├── profiles_regroup.go         # order-independent history pass
├── profiles_similarity.go      # classify on observation
├── profiles_learning.go        # filters, hierarchy
└── scheduler_controller.go     # non-holding pass, expiring identity cache

The classification has its own table because old clients share this database and reject unknown kinds in learning_similarity.

Evidence

Measured on a private copy of real history with the commit's binary. The live database was only read.

	Before	After
Profiles	1880	1433
Second preview		0 merges
Hierarchy rows		49 (48 projects + "Other projects" covering 72)
Client built from main		previews and lists the regrouped database

• Tests: core-1, core-2, ./internal/remote, test:web and lint all pass through the queue on 0b4cb02. The only commit after that adds the handoff doc.
• Property tests: 30 seeded shuffles of the metadata give identical classes and merge plans. A pass over a corrupt file, a corrupt sidecar and a 0644 file still completes.
• Mutations: order, alias fallback, registry-full skip, retry timing and failure holding each fail a test when broken, using go test -overlay.
• Reviews: independent Opus reviews drove the repairs. The last small fixes (blank-field ordering, retry test, shutdown log, docs) had focused tests and mutations but no further full review.

Merge Danger

Door: one-way for grouping on an upgraded machine

The pass merges profiles in the operator's history once. An operator can split a family back out, but there is no bulk undo. Pinned profiles are untouched and the new table is additive, so older clients keep working.

Blast Radius: scheduling estimates

• Over-budget predictions: 8 groups covering 424 families and 496 of 3,424 runs predict over the 13.6 GiB memory budget. Adaptive mode refuses these automatically until a variant has five successful runs of its own. Fixed and shadow modes are unaffected.
• Limited reduction: about 1,291 of the 1,433 groups still hold one family, because 1,421 of 2,722 families have no verifiable repository.
• Failure visibility: a permanently failing pass shows only in the resource controller log, because it no longer holds admissions.
• Approximations: older env and fnm jobs count as shell and javascript. A reused path can assign a wrong repository through a stored scope.
• Not diagnosed: "Cannot establish process identity" failed two tests once under load and passed on rerun. It comes from ownership.go, which this PR does not touch.

Full limits and next steps are in docs/handoffs/learning-grouping.md.

🤖 Generated with Claude Code