pgw#1009 (P1, catalog-blocking): the clone producer's checkpoint facts reach the publish declare

[PaulFidika]

Supersedes #526 (closed): that PR's branch carried id 1008, which the arm-gate
security lane claimed concurrently. Same commit, renumbered to pgw#1009.

The bug

th#1411 refuses a v2 publish into a repo whose live rows already carry
classification unless the request restates objective / distilled
(a v2 publish mints a new checkpoint identity and inherits nothing, th#1400).

run_clone accepts both facts (pgw#654) and spends them only on
apply_objective_scheduler_config — the scheduler stamp written inside the
produced tree. They never reach publish_v2, so the declare goes out bare.

publish_flavors (convert/publish.py) has passed both since pgw#654. This
call site is the one that did not, and git log -S "objective=objective_fact"
shows it never has.

Consequence

clone-huggingface cannot publish into any mirror the catalog already
serves. Not "sometimes" — every classified repo, every time, at declare,
before a byte moves. Mirrors are exactly the repos that get classified, so the
producer that creates them cannot update them.

Live cost — ie#609: master's tensorhub/wan22-t2v-a14b is missing
transformer_2/ (the mirror was split into source_include legs and the third
leg was never submitted). The repair is one more leg:

{"huggingface_repo": "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
 "source_revision": "5be7df96…",
 "destination_repo": "tensorhub/wan22-t2v-a14b",
 "source_include": ["model_index.json", "transformer_2/*"],
 "overwrite_repo": false, "destination_repo_tags": ["prod"],
 "objective": "flow", "distilled": false}

Refused twice on the master standing stack — 074a2c19-5988-409d-a09b-9362e90cacaa
(facts omitted) and 362957ad-ee57-4927-9611-ac9a04cc0bf0 (facts stated, same
refusal, which is what proved they never reach the wire) — both
400 classification_required … predecessor sha256:3b57e96996aa…. Every wan-2.2
function on the authoritative catalog is down until this ships.

The fix

Two kwargs on the publish_v2(...) call in clone.py, matching
publish_flavors. objective_fact / distilled_fact are the already
validated, already normalized locals.

Tests

tests/convert/test_clone_classification_pgw1009.py, hermetic against the
shared fake hub, driving the real run_clone:

• test_declared_facts_reach_the_publish_declare — the ie#609 repair shape
  (merge leg into a classified mirror): the declare carries
  objective: "flow", distilled: false, mode: "merge".
• test_undeclared_objective_stays_off_the_wire — an unstated objective is
  omitted rather than authored; distilled is a plain bool on the input
  struct, so an undeclared clone honestly states false.

Revert-turns-red: both fail with the two kwargs removed (verified).
Full tests/convert/: 74 passed, 2 skipped, 1 xfailed.

Deploy chain this unblocks

gen-worker release → training-endpoints/conversion pin (today
gen-worker==0.91.4; master's deployed conversion:prod 0.8.22 runs 0.91.4)
→ rebuild + redeploy conversion → re-run the ie#609 repair leg.

Tracker: pgw#1009, ie#609.

🤖 Generated with Claude Code

[PaulFidika]

Supporting evidence that this is an omission from the th#1411 sweep, not a design choice:

every other publishing function in training-endpoints/conversion carries the same
# pgw#654/th#1411 checkpoint-fact override; unset = publish_flavors … note and threads the
facts through — fuse.py, transform.py, quant/bnb.py, quant/modelopt.py, quant/svdq.py,
lora_card.py, prompt_corpus.py, _common.py. All of them publish through
publish_flavors, which was fixed.

mirror.py is the only one that publishes through run_clone, and it is the only one still
broken. It is also the only producer that ingests from outside the catalog, so there is no
other conversion function a caller can fall back to: clone-huggingface / clone-civitai are
the sole path from an upstream repo into a tensorhub mirror.

[PaulFidika]

Release vehicle — verified at source, so whoever merges this does not have to guess.

• v0.93.4 is already tagged at 34124789 and awaiting its CI-on-tag publish run.
  This commit is not an ancestor of it (git merge-base --is-ancestor → false), and
  34124789:src/gen_worker/convert/clone.py still has the bare publish_v2(...) call.
  PyPI's newest is 0.93.3. So this fix lands in 0.93.5 or later.
• The consumer that needs it is training-endpoints/conversion. te#167 PR B
  (167-repin-093-conversion) is already re-pointed at gen-worker 0.93.4 + conversion 0.8.29+.
  If PR B lands on 0.93.4 it will not carry this fix and a second conversion bump is owed.
• Master's deployed conversion:prod 0.8.22 runs 0.91.4 (read off the pod's
  worker_boot_phases.gen_worker_version, not off the checkout), so the endpoint is behind
  either way.
• te#167's own standing warning applies to the redeploy: do not deploy conversion mid-eval
  — check te#165's tracker tail for an in-flight slice first.

ie#609 (the master wan-2.2 outage) stays open until that chain completes. Its other three items
are done: the five wrong-tree bindings are rebound live on master, disk_gb is corrected
(inference-endpoints #267, merged), and the root cause is banked.