[rwkeyes]

Problem and result

Section 6 left ~165 ms of per-request fixed cost unnamed and said naming it needed a phase timer and a build. It did
not: engine 0.1.59 already prints the prompt path when STRATA_TRACE is set, and it names the missing phase.

real tokens	read	refill	lent slots	refill per slot
127	252 ms	71.7 ms	225	0.319
457	432 ms	82.6 ms	259	0.319
1,609	1,245 ms	188.2 ms	589	0.320

The expert-slot refill runs after the prompt read and before prompt done, so the first token waits for it — 72 ms
on a short prompt, 188 ms at 1.6 k — at a constant 0.319 ms per lent expert slot. The read itself carries 109 ms of
fixed work (109 ms + 0.706 ms/token). 109 + 72 + 56 ms of cache bookkeeping is the 165 ms.

That also reconciles §6 with the dispatch: pinning the cache on raises TTFT by 148–234 ms but raises read + refill by
only 7–29 ms, so the cache-on cost is checkpoint work inside the engine's prompt span, not in the read.

The lever: the refill does not have to be synchronous — it reclaims slots the next request lends again, and decode
needs its own experts. Moving it after the first token is 72–188 ms off TTFT for a few lines of engine change. It is now
item 1 of §4.

Related discussion

• [Docs] File the performance inventory and close the PLE read-path knobs #14 — the fixed-cost split this completes
• [Docs] Mixed traffic: keep batch_after at 4 (1 costs 1.9 s per conversation) #17 / [Docs] The 0-slot cache profile saves nothing and severs chat (retracts a 58 ms claim) #18 — the mixed-traffic and 0-slot results that fixed the measurement frame

Approach

1. Ran the ladder at 75 / 512 / 2,048 target tokens with STRATA_TRACE=1, six samples per size, and parsed the prompt
   path per request (strata-trace-split.py pairs each trace block with its prompt line).
2. Repeated the same window with conv_route: shadow to separate the cache-on cost from the read, which is how the
   §6 discrepancy was found rather than smoothed over.
3. Recorded the attempt on a second weight set (Q2_0) and its three refusals instead of dropping it.
4. The page gains §9 (the table, four findings, the lever, and the blocked cross-weight check); §6's item 6 now points
   at it; the ranked list gains the refill as item 1 and closes the phase-timer item as "done, no build".

Validation

Passed:

• 0.319 ms/slot at three sizes — 0.3187, 0.3189, 0.3196 — a rate that stable is a transfer, not noise, and it makes
  the mechanism checkable by anyone with the logs.
• The refill's position is read from the trace order, not inferred: read … 1246.9 ms → refilled 589 slots … 188.0 ms → prompt done, with the ladder's TTFT (1,463 ms at 1,610 tokens) matching read + refill + ~21 ms of server.
• The shadow arm reproduces the §6 gap in the right place: TTFT +148/+207/+234 ms against read+refill +7/+9/+29 ms.
• Every window restored production: :8095 http=200, engine 0.1.58, after each of the three runs.
• Artifacts complete: both engine logs with all trace lines, both ladders, the parser, the driver, and four Q2_0
  failure logs.

Failed:

• STRATA_TRACE prints no admission trace at --batch 0 (that block is batch-only), so the first attempt looked
  like the env had not reached the engine. The prompt path is what the question needed; the caveat is in §9.
• The Q2_0 cross-weight arm could not start — three refusals in sequence, each fixed before the next appeared:
  duplicate split shard number (PLE shard doubled as the sibling), then is a native (IQ) pack: it needs --native,
  then REFUSED: 11158 experts are neither in VRAM nor mirrored. Root cause of the last one: the Q2_0 pack declares
  48×512 experts where the shipped profile is 48×256, so it needs its own calibration.

Not run:

• No engine change: deferring the refill is proposed, not implemented or measured.
• No cross-weight confirmation, so every number here stands on one weight set (IQ1_M) — stated in §10.

Limits

• The unaccounted 56 ms is prompt-span minus read minus refill, which is subtraction, not an instrumented phase; it
  could hold more than one thing.
• One card, one model, one run per arm; the per-slot rate is measured at three prompt sizes, not across contexts.