sync(PINPORT): steps 5-7 of the cdefd9d499 cycle -- port the queue, re-verify, advance the pin

Row: -

Owed by .agents/upstream-sync.md §"The sync cycle", which owns steps 1-7.
#2490 tracked steps 1-4 and
delivered the report .agents/sync/2026-09-01-cdefd9d.md. This issue tracks
steps 5, 6 and 7: port the PORT-NOW queue, re-verify, and advance the pin.

Target. cdefd9d4997f00da72dc6245cc60678b50761b7e (2026-09-02) against the
active pin 5559679229bc961848b121ccdeaa8fa5d79bec98 (2026-07-26). The range is
1566 commits, 1077 of which touch the mirrored surface: 315 PORT-NOW, 356
INVENTORY, 406 IGNORE.

The pin has NOT advanced and must not advance until the queue is worked. A
pin advanced over an unported queue is worse than no advance, because every
downstream gate then measures against an oracle the tree does not mirror.

Owed

• The 315-item PORT-NOW queue. It is more than one wave. Take it in the
  dependency order named in the report's §11, not in SHA order.
• Disposition re-derivation. The report's own carry-over C7 records that no
  disposition was independently re-derived against this tree. First sampling
  suggests a substantial fraction of PORT-NOW entries are already satisfied by
  this tree's structure or are inapplicable, so 315 is an upper bound on work
  rather than a count of code gaps. The queue needs re-deriving before it is
  worked, or waves will spend their budget confirming non-gaps.
• Gateability of the target. gateable = no. Nobody has built
  cdefd9d499. AGENTS.md requires an oracle to demonstrably build and run the
  model before it can gate. FlashInfer moves 0.6.15.post1 -> 0.6.18 and the
  transformers floor 5.5.3 -> 5.10.4; torch is unchanged at 2.13.0.
  6a962071bd vllm#52998 flips VLLM_ALLREDUCE_USE_FLASHINFER on by default,
  which moves the vLLM production configuration used as the denominator.
• Re-measurement before comparison. Any benchmark baselined against the pin
  needs its vLLM denominator re-measured at the target (step 6).
• Tiering the 356 INVENTORY entries in .agents/porting-inventory.md.

Ordering constraints the report established

1. The KV cache layout refactor is a six-stage series. Stages 1-3 are at the
   pin; 4, 5 and 6 are all in range and must land in order: 61874f9842
   (vllm#51612) -> 57bd0ed441 (vllm#51704) -> 8bdc70ec7b (vllm#51718, the
   largest item at 1797 lines over 30 ported files) -> then e126687a9a
   (vllm#53896), which reworks the same kv_cache_utils.py surface.
2. Two revert pairs sit fully inside the range (47ececb58e/3d204dfdaa,
   bebf918044/9a4fd57cac). Porting either half alone lands a capability the
   target does not have.
3. 5789897aa4 (vllm#49969) adds a DSpark block-size validation that
   68ca6fd02c (vllm#50869) and 51def7848a (vllm#52223) remove. SHA-ordered
   porting installs it and then deletes it.

Deviations

Four recorded deviations in .agents/porting-inventory.md §9 conflict with this
advance, and the guide requires them updated in the same cycle. Three discharge
commits are identified: vllm#51655 -> 6adad08767, vllm#50210 -> ad5d29db70,
vllm#51255 -> 9035151d6c.

Start with .agents/upstream-sync.md §"The sync cycle", the report's §11, and .agents/porting-inventory.md. Re-derive the PORT-NOW queue in dependency order, account for the listed ordering constraints and deviations, and verify the target oracle and benchmark denominator. Done means the queue is addressed, inventory is tiered, the target is re-verified, and the pin advances only when the gates pass.