[joelteply]

tests: one learn fixture owns custody, grants and release for every learn script

Every learn now runs inside custody the core registers (weights, optimizer and workspace
residencies from the engine's own preflight) under a ledger grant, and every lifecycle verb names
the engine incarnation. Seven scripts predated that; each copied its own transport and none could
run. tools/server/tests/learn_fixture.py is the one owner:

• Engine(url, name): the incarnation, an idle-only grant, request/post/ok/get with the engine
  added on bound routes;
• fresh()/load(): dry-run plan, registered weights residency, seat;
• learn()/experience(): preflight THIS example, optimizer residency once per adapter, a
  workspace per example, split exactly as admit_learning_residency splits rows;
• shared_grant(measured, ...): a contended share whose latency allowance covers the measured
  chunks and boundary, max(entry, restoration);
• filling(): the longest reply the slot window holds, counted by the server's tokenizer;
• close(), also at exit: retire adapters, release every reference, revoke grants. Without it the
  engine-wide workspace refused the next script's owner (owner_mismatch, reproduced here).

Migrated: dream-adapter-routes.py (CI), learn-past-example.py, learn-unload-during-learn.py,
learn-every-format.py (all six modes), learn-cancel-custody.py, learn-chunk-yields.py,
learn-refuses-repacked.py. No script needs a hand-placed shadow-05.gguf any more: shadows are
minted fresh (two from one seed are the "copy" two-minds needs); the F16 copy is still made by
make-f16-shadow.py, from a fresh shadow.

Measured on a CPU llama-server built from this branch (IntelMac, stories260K, --no-repack):
dream-adapter-routes twice then learn-past-example twice on ONE server: pass (this failed with
owner_mismatch before close existed); learn-every-format default, --shadow-retirement,
--two-minds, --non-f32-refused: pass; learn-cancel-custody: pass; learn-refuses-repacked
(server without --no-repack): pass.

Not passing, reported, not hidden:

• learn-unload-during-learn.py aborts the engine 2 of 3 runs: ggml-opt.cpp:969
  GGML_ASSERT(!allocation_pending && !graph_pending) in ggml_opt_free <- opt_release, freeing idle
  B while A learns (confirmed from source by Codex; the native fix is assigned). This script is
  the regression for it.
• --fairness and learn-chunk-yields learn under contention: a contended graph chunk (1.5 s on
  chunk-yields) exceeds an allowance only idle steps can measure, so the step is refused mid-way
  (chunk_exceeds_latency). How a contended learn obtains its allowance is the ledger owner's open
  question.
• --template-replies needs a model whose template renders tool_calls; stories260K's does not.

Stacked on #106 (cormac/inherited-ci-repair); retarget to codex/s4-per-mind-integration after it lands. Card f1d9d2cf.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FRQWzgSfo79JtnwZywHKrE

[joelteply]

Review (Fable): approve. One owner of the learn lifecycle for every learn script (grants, custody registered from the engine's own dry-run rows, seating, release), closed on atexit so a failed assertion never leaves custody bound for the next script; seven scripts drop their hand-rolled copies (+419/-276).