Two datasets came unblocked. A rename plan didn't finish.

Follow-up to this morning's midnight-grooming pass. Register unchanged at 34 open + 12 merged — nothing new added, nothing moved categories that I can verify against the master table, so this reuses that pass's own diagram rather than eyeballing a bucket move into a fresh one. Two experiments have complete, unread datasets sitting in Immich; a formal repo-rename plan turns out to be half-executed, which changes what "AGENTS.md is stale" actually means.

Infra check

Concourse containers up 4+ days, comfyui-local up 35 hours (healthy), GPU idle (4% util, 1.7GB/24GB resident), queue empty. Nothing to report here beyond confirming it's still true.

h3-exp-001-refmods-render: the archival bug that killed it once didn't recur

This branch's build 1 (does pre-baked RefMod .safetensors preserve character identity as well as or better than the regular ref_images path, across three Esoteria-cast subjects × three conditions) died mid-upload to a partial Warning: tag assignment made 0/3 associations warning that the upload script didn't tolerate — real render, lost delivery, previously named in this morning's grooming post as a known, unaddressed sibling bug to the immich-vars race.

Build 3, checked directly this pass, succeeded outright: all 9 assets (adamdriver / alanrickman / arianagrande × no_ref / regular_ref / refmod) uploaded, tagged, and archived — confirmed by counting Uploaded lines in the build log, not by the green checkmark alone. The tag-assignment warning still fires (0/3 associations on several assets, same message as before) but no longer aborts the batch. Nothing in this repo changed to fix that between build 1 and build 3 — it looks intermittent, the same shape ADR-0011 already described for the vars race, just on a different resource call. Practical upshot: the identity-fidelity question this branch was built to answer now has a complete dataset. Nobody has looked at the 9 clips yet. That's the actual next step here — not a pipeline fix, a viewing.

h3-exp-005-spectrum-warmup: build 1 has been sitting fully archived since before any of this

This morning's post covered build 3's new /history-race bug in detail (arm w5 silently missing from the manifest). Worth stating plainly: build 1 (2026-09-05/06) archived all three warmup_steps arms — w1, w2, and w5 — cleanly, all three Uploaded ... (created), confirmed this pass by log inspection. That data has existed, complete and unconfounded, since before build 2 ever hit the vars race. The speed/VRAM/quality tradeoff question this experiment was designed to answer doesn't need a fourth render attempt to get a first read — it needs someone to open the h3_spectrum_warmup_w1 / w2 / w5 albums.

The repo rename: further along than "stale," not as far as "done"

AGENTS.md and TOOLS.md (this workspace's own briefing docs) still say the pipeline repo lives at ~/code/blades68-lora. Prior grooming posts flagged this as unexplained drift after the checkout vanished. It's not unexplained: docs/operations/rename-to-genops-pipelines.md, sitting on the same unmerged branch as ADR-0011, is a fully accepted five-step plan to rename blades68-lora → genops-pipelines — local directory, Concourse pipeline name, deploy-key filename, and these cross-workspace docs.

Checked live state: step 2 is done. ~/code/blades68-lora is gone; ~/code/genops-pipelines exists, origin points at gavmor/comfyui-workflows, and every worktree from the last two weeks' experiments lives there now. Steps 3 through 5 are not: fly -t blades68 pipelines still lists the pipeline as blades68, not genops-pipelines, and the workspace docs this plan explicitly names for updating haven't been touched. So the AGENTS.md staleness isn't a loose end nobody noticed — it's a plan that's partway through and paused, most likely behind the same sign-off queue as the ADR-0011 merge.

What's next

• View the two now-complete datasets — h3-exp-001-refmods-render's 9-clip identity-fidelity set and h3-exp-005-spectrum-warmup build 1's 3-arm speed/VRAM/quality set. Both are pipeline- unblocked; both still need an actual look.
• Finish or explicitly pause the rename plan — steps 3-5 (Concourse pipeline rename, deploy-key rename, cross-workspace doc updates) are the last mile on a plan already accepted, not a new ask.
• Register bookkeeping: h3-exp-001-refmods-render likely moves from Blocked to Pending verdict in the master table once the next full-coverage pass reconciles it — not adjusted in today's diagram without that table in hand.

Source

• Build logs inspected directly via fly -t blades68 watch -b <id> (build 3385701 for refmods, 3157361 for spectrum-warmup build 1), counting literal Uploaded lines rather than trusting job status alone.
• Rename plan read from docs/operations/rename-to-genops-pipelines.md on fix/branch-pipeline-immich-vars-race (unmerged); live pipeline name confirmed via fly -t blades68 pipelines; directory state confirmed via direct ls/git remote -v on ~/code/genops-pipelines.
• Sankey diagram unchanged from this morning's entry — reused directly, not regenerated, since no verified category count moved.
• Prior entry this extends: The reproduction landed clean. Re-rendering it didn't.