pgw#993: one flattening rule — carried_by resolves through the same expansion dynamic_shapes mirrors

[PaulFidika]

P0. Blocks pgw#868 A2 entirely — z-image is the only family in the fleet using the list-container vocabulary and, before this, it could never mint an AOT cell.

The defect

gen-worker 0.93.2, pod 7evvazd2xplzml (A100-SXM4-80GB), $0.4655: the mint passed every earlier gate, entered the AOT recipe, and refused in trace_graph — exit=2, deterministic, four identical attempts.

entry 'transformer/adapter=true,cfg=true': declared-range gate: declared dynamic dim
names input 'x', which is not a user input of the exported program
(inputs: ['cap_feats_0', 'cap_feats_1', 'lora_a', 'lora_b', 't', 'x_0', 'x_1'])

Two SDK features that each work and could not compose:

• Input.repeat containers (pgw#853) are flattened by torch.export into one positional user input per element, suffixed _0, _1, … for every arity — a one-element container is x_0, never x (re-measured here on a real export, both arms of the CFG fork).
• Dim.carried_by names its input by the declared name, and the declared-range gate resolved that name against the exported program.

So a Dim carried by a repeated container was unsatisfiable by construction. No declaration edit fixes it: naming x_0 breaks the cfg=false arm, dropping the dims forfeits dynamic-collapse — the reason z-image is 2 entries and not 20.

The fix — the invariant, not the patch

aot_mint.exported_input_names is now the ONE expansion rule. dynamic_shapes_spec mirrors the container structure with it; declared_range_gaps and lifted_input_gaps resolve declared names against exported ones with it. All three take the same per-arm arity map (aot_declaration.container_arities) the example feed was built from, computed once at the mint call site.

A gate that resolves declared names against exported names must share the name-mangling with the code that produced them; two independent spellings of one mapping is the defect class — which is why lifted_input_gaps is in this change rather than left for the next rented pod (pgw#993 acceptance item 4).

Every element is gated, not merely found: a per-element specialization is still refused, under the element's own name. Non-container declarations resolve exactly as before (pgw#846).

RED, off-GPU

tests/test_dim_flattening_pgw993.py drives the pod's program as a double (user inputs verbatim, refusal string verbatim), then exports for real on CPU for both arms. Reproduced on v0.93.2 and origin/master, green here:

N=1 user_inputs=['x_0', 't']         gate(declared names) -> the refusal above
N=2 user_inputs=['x_0', 'x_1', 't']  gate(declared names) -> the refusal above
N=1 / N=2 with the arity map         -> []

11 rows; mypy clean, ruff clean, lint_http_timeouts / lint_unreached_surface / lint_config_reads green; both suites run locally.

Also lands on 0.93.3

Per the tracker's release-vehicle rules, this lane cuts no branch and no tag: the mint-speed lane cherry-picks this commit onto release/0.93.3 (based on the v0.93.2 tag, pre-wire-v1) alongside pgw#992/pgw#989.

Not in scope, filed separately

The serve-side ingress contract has the same shape of gap one gate later: aot_package.input_contract records position as the index among exported (flattened) inputs while aot_serve.bind_call_inputs matches that position against the caller's pre-flattening args, so a container input binds the whole list to element 0 and shifts every later input. It is unreachable until a container family actually mints — i.e. now. Reported to the tracker rather than folded in here, because it needs its own design (element identity in the contract) and its own RED.