Ghost calibration MEASURED on hardware: straggler 1.462x → 1.189x

[defenwycke]

Been sitting unmerged on measure/ghost-calibration since 2026-09-05. Raising it because the result is real, and because docs/BUILDS.md §3 currently states the opposite.

What it measures

Two L40S, block 962,000, 16 chunks, real GPU proving. Same guest, same block, one box per arm, differing only in the four HAZYNC_COST_* values:

arm	straggler	max	mean
shipped defaults	1.462x	149 s	102 s
refit	1.189x	124 s	104 s

Ghost's own constants, fitted on block 965,500 (7.7% Schnorr), intercept forced to zero:

HAZYNC_COST_EC_OP=85636  HAZYNC_COST_SCHNORR_OP=168542
HAZYNC_COST_INPUT_BYTE=2 HAZYNC_COST_INPUT_BASE=53162     Schnorr:ECDSA = 1.97x

⚠ The card arithmetic was NOT re-derived. Read this as "the ≤1.35 straggler target is met", not as a card count.

⛔ This corrects a live claim in BUILDS.md

§3 says: "Ghost needs its own calibration; it does not have one." It does — this is it. That line was written before this run and never updated, and it has already been quoted as the measured reason Ghost cannot be a shipping channel (hazync#225).

The channel decision does not change, but the honest reason does:

• ⛔ not "Ghost is uncalibrated" — it is calibrated, and its refit straggler 1.189 is better than Core's 1.210
• ✅ Ghost is substitution (broad) — ECDSA and Schnorr group arithmetic, scalar inverse and the SHA fastpath all replaced. That is the fidelity call, and it stands on its own without needing a performance excuse.

⚠ Note what does NOT transfer: Core's refit and Ghost's are different constant sets. Applying Core's to Ghost moves its straggler 1.407 → 1.884. Each channel carries its own.

⏰ The part worth keeping

This run first measured 1.567x — worse than doing nothing — and that was reported. It was one chunk. Across 32 chunk timings every sample sits at 2.4–3.0 s/segment; the offending chunk came in at 4.89 s/segment and never reproduced (re-run twice: 94 s, 94 s, 2.69 s/seg).

Two of the three checks that recovered it refuted a hypothesis rather than confirming one:

1. The segment straggler said the opposite (57 → 49 max) — the machine-independent metric is what flagged the wall-clock as wrong.
2. "The boxes differ in speed" — plausible and false: identical 33-segment chunks run 87 s vs 90 s, and a straggler is a ratio, so box speed cancels out entirely.
3. Wall time is proportional to segments across 31 of 32 samples, isolating the 32nd as a transient.

⇒ A single un-replicated chunk timing was enough to invert the conclusion. Re-run an outlier before believing it, and prefer the quantised metric (segments) over the continuous one (cycles) — proving bills in whole segments.