pgw#999 (P0): the classified adopt refusal reaches the wire, and the rig covers the gate that produced it

[PaulFidika]

Attempt 26's sdxl mint succeeded — 36/36 entries exported, compiled, sealed, finalized — and was then refused by the runtime that had just built it. Three abort events fired and not one named a cause. 2 h 45 m and $2.72 of L40S bought the word "something", and the pod is gone.

The discard, exactly

provision.arm_aot returns a classified AdoptOutcome; aot_serve raises AdoptError with a reason token. fleet_cells.adopt_delegated_mint spent both on:

armed = bool(provision.arm_aot(...))

…and its except Exception branch wrote the reason to a logger no pod exposes. That is the worker-errors-to-orchestrator defect class verbatim.

The fix

phase is the countable column, so it carries the class — the convention self_mint_skipped already uses — and the detail quotes the gate's own sentence. Every branch classifies:

branch	phase
returned miss	its own reason (contract_invalid, no_arm_for_mode, numerics_refused, …)
raised AdoptError	.reason (constants_constant_unresolved, …)
raised anything else	the exception TYPE, not a flattened token
CellSelectionBugError	cell_selection_bug (th#883's loud class, kept separate)
falsy arm, no classification	unclassified_arm_refusal — never a blank phase, which reads as "no reason exists"

The classification is recorded on the pending, so DelegatedResult.reason and the executor's terminal error quote the one string produced where it was known, instead of three vocabularies for one fact.

Nine tests, all RED at HEAD, all GREEN here — verified by running them against the unpatched sources, not assumed.

The rig gap this found

task rig:micro's parity leg armed through aot_serve.enable. The delegated mint arms through provision.arm_aot, which additionally does the mode route, the lifted-LoRA install for a bucket-bearing cfg, and the numerics gate. The green cycle reported yesterday did not cover the path that refused sdxl. It does now, and stays green: 26.0 s, three entries, parity 7.15e-07.

Diagnosis so far — no pod, real path

Reproduced end to end through the real adopt_delegated_mint: a parent whose module does not expose an FQN the child traced is refused constants_constant_unresolved, naming the FQN (proj_in.weight (source=state_dict)).

Two hypotheses ruled out, not merely suspected:

• extra per-layer buffers are harmless to constant binding (measured — a parent with +16 weight_scale buffers adopts cleanly);
• the SDK's w8a8 Linear preserves the weight FQN (register_buffer("weight", …)) and only adds weight_scale/input_scale, so sdxl's quantized lane does not refuse by a renamed weight.

lane_unavailable is not reachable from this path at all: verify_declared gates on format, kind, package_constants_in_so, sm/torch/cuda, host_isa and family — and on no execution lane anywhere.