EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a gotcha_candidate: yes failure-lesson raised on task #1315 (emitting agent: experiment-implementer, crash-fix round 8).

Goal

Extend upload-policy.md's 429/no-path material: a dispatcher seam that FAIL-FASTS on hub._upload's no-path return must wrap it in a bounded jittered-backoff retry (transport-class only, same fail-loud error on exhaustion) — the complement of the existing #488 rule (which bans the opposite failure mode, silent warning-and-continue).

Workflow gap

Bug observed: #1315 r8: two p11_upload kills ~35 min apart — hub._upload logged 'Upload failed: 429' / Xet 'maximum queue size reached', returned no path, and the dispatcher's fail-fast _up converted the retriable rate limit into a run-killing final-phase crash twice.
Why it is a workflow gap: upload-policy.md's existing "upload returned no path" text (the #488 entry, lines ~243-252) covers ONLY the silent-loss direction (no-path must be a TRACKED GAP, never warning-and-continue) and the sweep bulk-commit rule; it is silent on the fail-fast direction — a seam that correctly refuses to continue still needs a bounded transport retry BEFORE its fail-loud raise, or sustained fleet HF traffic kills runs at their final phase.
Confidence (emitter): high (root_cause_confirmed: yes; fix + retry tests landed on issue-1315 @ c3c600541f)
verified-at-filing: grep -ciE "upload.wedge|429" .claude/rules/upload-policy.md → 5 hits (presence of the 429/wedge section confirmed); grep -niE "no.path|returned no path" .claude/rules/upload-policy.md → 2 hits at lines 245/250, BOTH the #488 silent-loss rule (inspected: they mandate tracked-gap-never-continue and say nothing about bounded retry at a fail-fast seam) — the absence-of-retry-guidance claim binds (2026-07-15)

Proposed change (candidate diff sketch — refine in planning)

In .claude/rules/upload-policy.md, adjacent to the #488 "upload returned no path" rule:

(c) the FAIL-FAST direction needs a bounded retry: a dispatcher seam that
raises on the no-path return (correct — never warning-and-continue) must
first retry the transport class (429 / 5xx / Xet queue-full surface as a
no-path return) with jittered exponential backoff (~3 attempts,
30/60/120s), then raise the SAME fail-loud error on exhaustion. Uploads
are idempotent (already-landed files skip-and-verify), so retries are
free; without them sustained fleet HF traffic kills runs at their final
upload phase (#1315: two p11 kills ~35 min apart). Worked example:
_upload_with_transport_retry() in scripts/issue1315_dispatch.py.

Scope / surfaces

Primary target: .claude/rules/upload-policy.md
Grep before editing (grep -rln 'returned no path' .claude/ CLAUDE.md scripts/); keep consistent with the #488 entry and the upload-wedge ladder; list hits in the plan.

Constraints / invariants

Workflow-surface only. scripts/workflow_lint.py no-flags run passes.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

workflow_fix_target: .claude/rules/upload-policy.md
fingerprint: 17023149c3b7

failure_class: infra phase: p11_upload (scripts/issue1315_dispatch.py) lesson: orchestrate.hub._upload swallows HF-429/Xet-queue transport failures and returns "" — a dispatcher seam that fail-fasts on the no-path return converts a retriable rate limit into a run-killing crash at the FINAL phase. Wrap the no-path return in a bounded backoff retry at the dispatcher seam (raise the same fail-loud error on exhaustion); uploads are idempotent so retries are free. generalizes: yes owning_agent: experiment-implementer gotcha_candidate: yes root_cause_confirmed: yes