fix: post-merge review sweep — lint blind spots, bagel prefill guard, label workflows, refl nits

[haonan3]

Summary

Fix batch from an adversarial post-merge review of the five recently merged PRs (#210, #277, #279, #284, #286). Every item below was reproduced/verified before fixing; the review evidence lives in the per-item notes.

Boundary lint had systematic blind spots (#279), and they already mattered on the tree:

• The cross-package rule missed from experimental import <sibling>, bare import experimental, relative imports climbing past the package root, and treated top-level experimental/*.py (incl. __init__.py) as a lint-free shared space — the most natural spellings of exactly the sideways borrowing the README forbids. All caught now, in any spelling.
• The "core dependency" set was [project.dependencies] only, so every extras pin — torch, torchvision, sglang, vllm — was re-declarable undetected; the in-tree videoalign/requirements.txt was already re-declaring torchvision with no signal. The set now unions base deps, every [project.optional-dependencies] extra, and [tool.uv] override-dependencies; the scan covers requirements*.txt and fails closed on pip option/include lines and URL/VCS lines it cannot vet. The videoalign re-pin is dropped (comment documents that the engine extra provides it) and the face extras file documents the same for its top-level torchvision import.
• New tier-direction rule in check_recipe_targets.py: recipes outside experimental/ may not _target_ into experimental.*, and a package's recipes may not target a sibling package — _target_ is this repo's primary wiring mechanism and previously carried zero directional signal.
• experimental/README.md now states which rules are script-enforced (1, 2, 4) vs review-enforced (3, 5, 6), and that private packages are lint-gated locally only.

Bagel (#277 follow-up): BagelPipeline._build_contexts runs the same vendored inference-signature prefills the #277 fix declared eval()-load-bearing, but with no guard — SFT/trainside callers reaching it in train() get the exact packed_query_sequence TypeError back. Same guard-and-restore applied.

#284 follow-ups: the two moved converters landed inside gitignored datasets/* (tracked but invisible to ruff/rg/editors; new sibling files needed add -f) — the dirs are whitelisted with their jsonl outputs kept local-only. Two recipe comments pointing at nonexistent scripts/*.sh launchers now point at examples/run_experiment_single_node.sh. experimental/private_* also covers stray private files now (was dir-only).

Label workflows (#286 follow-ups): merged/closed PRs shed their status label (they kept need review forever — see #210/#279 right now); synchronize self-heals a PR that lost its label to a missed event/transient failure; the review-labels consumer drops its shared concurrency group (GitHub keeps one pending run per group, so reviews landing on two PRs within seconds silently dropped the earlier sync).

refl nits (#210 follow-ups): seed: null no longer TypeErrors at the worker (int(params.seed) had no None branch; generate_latents already supports unseeded); the unreachable grad-input mismatch guard in worker.py is dropped (its dict-index comprehension KeyErrors first, and resolve() consumes the same keys earlier still).

Related Issue

N/A

Test Plan

• Negative/positive battery for both guards (34 checks, all pass): every blind spot above rebuilt as a synthetic violation in a lab tree — from-import/bare/relative/__init__-hub sideways imports flagged; extras pin, uv-override pin, case-variant name, -r include, VCS line, and requirements-gpu.txt coverage flagged; self-imports (absolute/relative/from-bare), core imports, additive names, and name @ url direct refs stay clean; core→experimental and sibling _target_s flagged while own-package recipes pass with no unresolved noise.
• Real tree: python lint/check_experimental_boundaries.py → ok; python lint/check_recipe_targets.py → 2310 recipe _target_ paths resolve. (unchanged count).
• git check-ignore battery: both converters no longer ignored; their *.jsonl outputs, other datasets/* dirs, experimental/private_notes.py, and experimental/private_a/f.py all ignored.
• python -m py_compile on all six edited .py files; bash -n on both edited workflow scripts (extracted from YAML).
• SKIP=no-commit-to-branch pre-commit run --all-files → all 16 hooks pass (both hardened guards included).
• Not run: GPU smoke for the bagel guard path (CPU box) — the change is mode-neutral by construction (guard-and-restore around an already-no_grad block, same pattern as the merged fix(bagel): rebuild replay KV contexts in eval() so navit keeps inference dispatch #277 fix); and the master_dtype: fp32 verification gap flagged for feat(experimental): add WAN ReFL as a self-contained training package #210 is deliberately NOT addressed here (owner action on the pinned-torch env, see Reviewer Notes).

Compatibility / Risk

• The boundary/targets guards get strictly stricter; the current tree passes clean. Future violations that previously slipped through will now fail pre-commit/CI — that is the point.
• Label workflows: behavior changes are label-lifecycle-only (closed cleanup, self-heal on synchronize, no shared concurrency). The pull_request_target consumers still check out nothing and interpolate no PR-controlled strings; the PR-number regex guard in the review consumer is untouched.
• Bagel: guard-and-restore around prefills; freeze/rollout paths unaffected (colocate short-circuits before this call).
• refl: seed: null now takes the supported unseeded path instead of crashing.

Reviewer Notes

• AI-assisted (Claude Code): five parallel review agents (one per merged PR) + orchestrator hand-verification of every HIGH/MED finding before fixing; the submitter reviewed the full diff. Duplicate-work check: no open PR touches these files (docs(algorithms): placement policy — core canon vs experimental variants (RFC) #280 touches only an algorithms README, feat(experimental): port SD3 image ReFL and remove the legacy core ReFL path #281 touches refl but none of these lines — the wan21/wan22 seed lines may need a trivial rebase there).
• Deliberately not in this PR: (1) feat(experimental): add WAN ReFL as a self-contained training package #210's master_dtype: fp32 default was never GPU-verified at the merged head (the recorded smoke ran torch 2.7.1 with fp32 dropped; the restore commit postdates it and the yaml itself says it needs the pinned torch ≥2.11) — needs an owner smoke on the pinned env + README table update; (2) narrowing the label workflows' issues: write — likely redundant per the 403 evidence, but should be verified on a real labeled run after this merges rather than narrowed blind; (3) a backfill dispatch (gh workflow run pr-review-status-labels.yml) to reconcile the stale labels fix(ci): restore pull-requests write for PR label workflows #286's test plan promised — one manual command, listed for the maintainer.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.

[haonan3]

bca7e1e8 addresses an external (Codex) review round — all three findings verified real and fixed:

1. Same-PR label race (P2): removing the global concurrency group traded cross-PR pending-slot drops for same-PR parallel runs racing label writes (and set -e dying on the 404 a parallel DELETE leaves). Now: per-PR group pr-review-status-labels-${{ workflow_run.pull_requests[0].number || run_id }} — same-repo PRs get back the serialize-latest semantics, fork PRs (empty pull_requests) fall back to per-run with a tolerant DELETE (|| true) absorbing the benign 404.
2. Tier-root recipe bypass (P2): experimental/root_escape.yaml targeting experimental.refl.* slipped through the len(parts) > 2 sibling guard. A recipe at the experimental/ root now errors — a recipe belongs inside one package.
3. No-space PEP 508 direct refs (P3): name@git+https://... was mis-flagged unparseable (only " @ " was recognized). @ joined the follower set — the name still gets extracted and checked against the core set (transformers@git+... correctly reports as a re-declared core dep), while bare git+/URL lines still fail closed. Kept the stdlib-only parser rather than importing packaging.

All three reproduced as lab cases in the negative battery before fixing (39 checks, all pass); full pre-commit run --all-files green.