release-train: develop -> staging

[LukasWodka]

Automated promotion by the release train (RFC-0008 D14). Head is the train-managed release-train/to-staging branch (a mirror of develop), so it never collides with a human PR. Merged only when the fr-gate is green.

---

Note

Low Risk
Comment-only changes with no runtime or CLI behavior changes.

Overview
Documentation-only update to internal/resources/resources.go comments so they match how client-runtime actually applies training resource envelopes.

The package header now states that GPU pods use the per-rank base multiplied by nvidia.com/gpu once bounded (with BestEffort behavior until then), and that the documented value is the CPU-pod ceiling.

The DefaultTraining comment is expanded to spell out per-rank base vs runtime scaling: CPU pods carry the value as-is; GPU pods get base × GPU count (with packing caps and admission rules); in-pod CPU is still split across ranks via engine thread-pool sizing. It also clarifies that tracebloc resources show reports the per-rank base, not the stamped multi-GPU envelope.

^{Reviewed by Cursor Bugbot for commit e812c5a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[LukasWodka]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit e812c5a. Configure here.}

[LukasWodka]

Functional review — staging batch, 2026-09-09 (FR reviewer: LukasWodka; self-signoff per RFC-BACKEND-1405 D6)

Decision: PASS. Every card this promotion carried that sat at FR on staging is advanced to Ready for prod. Evidence pulled, not reconstructed (README → Functional review):

pass	run	verdict
fr-assist (artifact channels)	release-train 34360917649	all OK — GH releases rc/latest (cli, client, data-ingestors), helm index stable-only, TestPyPI rc present and not behind prod, npm @latest stable / @staging newest rc, rc tags anchored to staging, cli smoke battery (installer + rc binary). 1 SKIPPED: cluster-scoped cli smoke (needs workspace credentials) — an absence, not a pass.
journey-tier-a, baseline none cell, staging harness against staging	e2e-test-agent 34368861675	success, incl. D10 · change of state
journey-fleet scope=tasks, staging harness against staging, one leg per task	e2e-test-agent 34368865960	12/16 green: causal_language_modeling, embeddings, image_classification, masked_language_modeling, object_detection, semantic_segmentation, sentence_pair_classification, tabular_classification, tabular_regression, text_classification, time_series_classification, token_classification. 4 red at Train a model on the use case: keypoint_detection, seq2seq, time_series_forecasting, time_to_event_prediction (BadRequest).
frontend-app smoke (Cypress, staging ref)	frontend-app 34369790687	success (builds staging's frontend code; the smoke targets the dev API by design)
staging deployment liveness	https://stg.tracebloc.io → 200 (/explore), https://stg-api.tracebloc.io → 401 (up, auth-gated)	up

The 4 red fleet legs are not staging regressions. The same four tasks are red on develop in today's fleet run 34343444708 (which additionally fails embeddings — green on staging) and in yesterday's proof run 34254097123. They are the open Family C defects tracked under backend#3289 (time_to_event_prediction → backend#3292; time_series_forecasting → backend#3294; seq2seq / keypoint_detection → backend#3291/#3297 lineage). Staging is at least as healthy as develop on every task.

What this evidence does not cover, per the README, and is advanced anyway on the operator's (Lukas's) explicit instruction to clear the FR queue for a prod hop: interactive/TTY CLI prompt behaviour; per-ticket behaviour (one journey verdict is shared by every card in the batch); design-system-v2 visual output beyond Tests green and the npm @staging rc publish (Chromatic is unarmed); any step after the failing train step in the four red tasks. Cards whose behaviour lives only there carry no automated evidence — the sign-off on them is a judgement, recorded here as such.

Performed by the release-train operator session; mirror→card mapping derived from /commits/{sha}/pulls over this PR's commits (not from subject text).