pgw#997: a REAL micro endpoint family — 3 export entries, a 15-second full mint cycle

[PaulFidika]

Closes the "sdxl is the test vehicle" problem: 6.9 GB, 36 export entries, ~95 min/pod cycle. This adds a real org worker with 3 entries whose full local machinery cycle is 15 seconds.

What's here

examples/micro-diffusion/ — a real org worker, not a fixture. Own pyproject.toml, endpoint.toml, Dockerfile-first build contract (gen_worker.discovery at build, gen_worker.entrypoint as ENTRYPOINT), catalog Slot(..., selected_by="model") with no default_checkpoint (pgw#969's shape), a registered export declaration, a @family defaults vocabulary, two served arms.

denoiser/cfg=true     guided arm; container arity 2
denoiser/cfg=false    turbo arm; container arity 1
decoder               a SECOND target, its own dims, no fork

Smallest set that still exercises plan selection, a fork coordinate, a derived dynamic range, a second target, entry naming, the seal, the publish wire and the cross-process adopt filter.

Two declared facts keep pgw#993/pgw#994 under test on every cycle:

• a repeat= CONTAINER input with a plain input immediately after it — t's argument position is 1, its FLATTENED position is 2. Asserted structurally in CI.
• one Dim carried by both a container element axis (("x", 0)) and a plain tensor axis (("latent", 1)).

Weights are generated, never fetched. (seed, config) -> bytes; nothing in git, nothing on a developer's box. The Dockerfile runs --verify (regenerate + byte-compare), so an image whose weights did not reproduce fails the build. 1.1 MB.

The rig is vehicle-selectable. --vehicle {tiny,micro} / task rig:micro; endpoint coupling moved to tests/harness/rig_vehicles.py. task rig:mint unchanged.

Measured (this box, CPU, synthetic sm per pgw#983)

leg	s
gates	0.9
weights (generated)	0.1
handoff (3 declared entries)	0.03
mint-child (real spawn, export + AOTI, seal)	9.4-10.6
publish (real CellPublisher, 4 HTTP calls)	0.2
adopt + parity (SECOND process)	2.7-4.3
full cycle	14.8-15.4

Adopt is a real parity check, not a filter check: served_entry_calls {decoder: 1, denoiser/cfg=false: 1, denoiser/cfg=true: 1}, execution_count 3, ingress_refusals 0, max |delta| vs eager 7.15e-07. That call is exactly where pgw#994 lives.

task rig:mint (the pgw#978 vehicle) still green: 18.6 s.

Found by the first cycle: pgw#998

unexpected None!, 512*s18*s57 — bisected to the mint's own torch.export.save/load hand-off to its compile child, which makes a nonlinear symbolic extent unlowerable. Not strictness, not the code-only config, not the wrapper split: the round trip alone. A squared symbol fails identically, so the trigger is nonlinearity, not symbol count. z-image's declaration has the same two-dynamic-axis shape and is worth checking before its next buy. Filed as pgw#998; this declaration sidesteps it with token sequences (flux/qwen's real interface) and a CI row keeps the invariant RED-able.

Not run

The pod runbook lives in examples/micro-diffusion/README.md, end to end, with derived ESTIMATEs (~5-8 min/cycle vs sdxl's ~95). No pod has been bought and none may be without Paul's go.

Coordination

Touches scripts/micro_mint_rig.py, which pgw#995's env-delivery lane also edits — hunks are disjoint (vehicle seam + adopt call here; docstring, run_cycle signature and child-env block there); whichever lands second rebases. pgw#996 touches aot_mint.py/export_contract.py/discover.py — no overlap.