Port upstream PR #6 (imperatormk) - fixes MIL scalar type syntax
from M4-only shorthand to canonical verbose format that compiles
on all Apple Silicon (M1/M2/M3/M4).

Changes:
- program(1.3) to program(1.0), ios18 to ios16 target
- Scalar type shorthand to canonical verbose format
- Simplified buildInfo dict (no M4-specific version strings)
- fp16 I/O fallback: g_fp16_io flag with auto-retry on compile
  failure for M1/M2 where cast op is unsupported
- Dynamic IOSurface byte calculation (bpe: 2 for fp16, 4 for fp32)

Tested on M1 Pro, macOS 26.3 (per upstream PR author).

…program(1.0), ios16 target, tensor types across 18 files

… hardening

#1 /models bypassed safety: now uses services.safe_compile (lint → validate → passes → compile)
instead of calling bridge.compile directly. The shared compile service enforces all /raw/kernels
safety rules (MIL lint, positive tensor-size validation, tensor-count limits, compiler passes,
weight-name validation, bridge-error mapping).

#2 Model leases leaked: register_model now does atomic replace (free old lease before installing
new). unregister_model calls bridge.free(kernel_id) before removing the dictionary entry.

maderix#3 Async eval bypassed admission: now calls admit_eval BEFORE dispatching the background task,
and finish_eval in _bg's finally block. max_pending backpressure applies to async work.

maderix#4 Cancellation resurrection: _bg checks sub.state == CANCELLED before transitioning to EXECUTING.
A cancelled submission is not resurrected.

maderix#5 Submission memory leak: SubmissionRegistry now has max_retained=100 + ttl_s=300. _prune()
evicts terminal submissions by TTL and count on every create().

maderix#6 Batch-eval bypassed admission: each item now goes through admit_eval/finish_eval. Endpoint
honestly named "serial batch" (NOT parallel without kernel replication).

138 tests.

Co-Authored-By: Claude <noreply@anthropic.com>

…models admission

#2 W&B: _process now handles event/run_finish (previously discarded). Fallback delegates
correctly when wandb absent (metrics + traces + events forwarded). finish_run finalizes in
queue order (no race). flush() drains with 5s timeout.
maderix#3 LocalJSON: writer calls task_done() (flush no longer deadlocks).
maderix#4 Backend registry: constructed in lifespan (ANEBackend + MLXBackend), stored on
app.state.backends.
maderix#6 /models: run_model now goes through admit_eval/finish_eval (previously bypassed
scheduler). register_model uses safe int parsing (try/except → 400, not 500).
maderix#8 /v1/resources: uses app.state.settings (was always false), ANE_PUBLIC_ENDPOINT env
(was guessing hostname), correct tailscale_serve detection. Lifespan sets
app.state.settings.
Lifespan flushes all observability backends on shutdown.
mlx has platform marker in requirements.txt (Darwin arm64 only).

149 tests.

Co-Authored-By: Claude <noreply@anthropic.com>

…, W&B fallback, deadline parsing

#1 All integration routes (/mlx/compare, /mlx/diff, /pipeline) now go through admit_eval /
finish_eval — no route bypasses the scheduler. Pipeline ANE steps also admit. The /mlx/compare
demo uses try/finally to guarantee kernel lease release.

#2 Batch-eval validates inputs_per_item >= 1 and rejects input counts not exactly divisible
(no silent tensor discard). Both deadline_ms parses (eval + batch) now use try/except → 400
for non-numeric X-Deadline-Ms (was 500).

maderix#3 W&B fallback: when wandb is not installed, log_metrics / log_event / log_trace delegate
directly to the fallback instead of queueing into an unconsumed queue (previous code started
no uploader thread without wandb → events stuck forever).

maderix#4 SubmissionRegistry.mark_terminal() calls _prune() immediately when a submission completes
(no waiting for the next create()).

maderix#6 test_models_register_rejected_by_lint: skips when ANE bridge unavailable (was asserting
422 on non-ANE hosts where /models returns 503 before reaching lint).

Pipeline spec JSON errors now return 400 (was 500). 149 tests.

Co-Authored-By: Claude <noreply@anthropic.com>

… mx.distributed, full output descriptors

maderix#6 Remote store qualified references: put/get/list/delete now use tier-prefixed IDs
(remote:<host>:<tid> or local:<tid>). A failed remote put that falls back to local returns
a local: ID; get/list/delete route to the correct tier based on the prefix. list() merges
both tiers. The namespace is coherent — no silent namespace collision between machines.

maderix#7 Deleted stale mx.distributed path: removed ane_distributed.py, mlx_worker.py, and
test_distributed.py. mx.distributed only supports single-machine multi-core; the TCP worker
(mlx_tcp_worker.py + ane_tcp_transport.py) is the real cross-machine path. The distributed
endpoints were already removed in the prior commit; now the dead code is gone.

maderix#8 ANE backend output descriptors: load() now stores full output_descriptors from the
ArtifactDescriptor spec (dtype + shape + byte_length) per executable_id. execute() constructs
output TensorDescriptors from the stored specs, not invented [1,N,1,1] shapes. unload() clears
the specs. Callers who provide output_descriptors in the artifact spec get exact output
metadata; callers who don't get byte_length-correct descriptors with empty shape.

7 tests (updated for qualified refs + merged list + fallback). 156 total.

Co-Authored-By: Claude <noreply@anthropic.com>"
git push -q origin ane-compute-api && echo "pushed"; git rev-list --count origin/main..HEAD