fix(patches): integrate graph-capture corruption guard on current main

[plotarmordev]

Summary

Current-main integration of @de1tydev's cross-repository PR #116. The original author and commit attribution are preserved; this branch exists because #116 predates merged PRs #112/#113 and cannot be updated in its fork without rewriting contributor history.

• remove the obsolete #50004 adaptive-topk backport from the pinned pre-#50004 image
• add upstream vLLM #52492's CUDA-capture guard to the #49486 shortcut
• preserve current main's HEADLESS healthcheck/CI gate and scripts/loop_detector.py
• record upstream #52823's later capture-safe re-land so the provenance statement remains current
• remove the obsolete #50004-only benchmark field and clarify the remaining counterfactual buffer comment

Supersedes #116 for merge. Closes the integration blocker found in independent review; no changes to the contributor fork.

Verification

• bash scripts/ci-validate.sh — PASS at 7f5538b90d5932f2fe78dc261a776ae6737d8c3d
• scripts/test-hotfix-atomic-transaction.py — 23/23 PASS
• all changed shell parsed; Compose rendered
• primary-source audit: #52492 guard byte-identical; #50004 removal is the inverse of upstream #51318 for this pinned image
• per-commit secret/source-identifier scan — 0 findings

Live two-rank recreate verification is in progress and will be posted before merge.

[plotarmordev]

Independent live validation at exact head 7f5538b90d5932f2fe78dc261a776ae6737d8c3d on the two-rank DGX Spark pair:

• stopped the active project and recreated both candidate containers from an isolated recipe/worker directory (not a container restart)
• head runtime: active_topk_width=0; one [PORT #52492] marker; hotfix-dsv4-skip-topk-49486.sh --status reports import/kernel/#49486/#52492 all APPLIED
• worker runtime: identical active_topk_width=0, capture marker count 1, and all four status checks APPLIED
• 4-way live chat smoke: 4/4 passed
• RULER-lite at 8K/32K × {S-NIAH, variable tracking}: 4/4 passed in 55 s
• served model remained deepseek-v4-flash-abliterated, max length 1,048,576

This validates the required container-recreate semantics and both-rank patch state. It does not claim a short run can prove the absence of an hours-later corruption event; the mechanism/source fix and exact runtime state are the evidence for that risk.

[plotarmordev]

Final combined-candidate evidence (e3d4bad7450c30ed9b5507a5f56aaf5d52121ec3, tree e26b6b2281ded34d9a7b247233e59ec362d3274e): both TP ranks report #49486 early return and #52492 capture guard APPLIED; #50004 is absent from the synchronized boot chain. Fresh-cache boot, C4 smoke 4/4, and repeated decode sentinels completed without target-kernel JIT, NCCL watchdog, or traceback lines after warmup. Independent CYBERGPT exact-head review remains APPROVE.