GPU experiment runbook: everything the free tier could not settle

[defenwycke]

Everything achievable without a card is now done — Tier 0 (E1–E4) measured, E5's chunk half
measured, E10 closed, E11's free half measured. This collects what's left, ordered so nothing
gets wasted.

The ordering is the point

1. Prove 16 chunks first. Three later experiments consume the receipts, and regenerating them
   is the expensive part.
2. E6 (processes × po2) before E9. E6 decides whether E9 is worth attempting: if worker
   processes fill the 65% GPU idle, E9 is dead; if VRAM caps them at one process at po2 22, only
   pipelining can fill it.
3. E7/E8 — po2 23 with 🚧 Vendor risc0-sys for upstream #3781's u32 overflow fix (needs B200 validation) #182's patch, newer driver. Take a po2 22 control on the same binary
   first, or the result means nothing.
4. E5 remainder, E11 GPU half, then E9 — in that order of cost.

E11's free half came back clean, and isolates the GPU question

quantity	N=4	N=16	invariant?
total execute cycles	14,341,307,904	14,341,307,904	exactly (also N=128)
total segments	3,411	~3,404	yes (0.2%)
total wire bytes	949.9 MB	~949 MB	yes
straggler	1.00×	1.00×	1.00× through N=64

Chunking is free all the way to the aggregate. Segments per chunk halve almost exactly on every
doubling (805 → 426 → 213 → 104 → 49), and HAZYNC_CHUNKS=16 is confirmed as a default nobody
revisited — the code says so outright.

That leaves E11's GPU half with a single unknown: does assumption resolution cost grow
sublinearly in N? If so, more chunks is close to free throughput and 16 is leaving parallelism
unused.

⚠ The N=4 total is measured across all four chunks. Extrapolating from chunk 0 gives 3,220 and
is wrong — chunk 0 is the smallest (805 against 864/871/871).

Guards, because each caught a real false result today

• Read METHOD_ID with host method-id, never by scraping the binary —
  strings | grep -oE '[0-9a-f]{64}' returns 000000000019d668…, the Bitcoin genesis hash,
  which is 64 hex characters and entirely plausible.
• Record METHOD_ID per arm. A build that silently didn't happen returns identical numbers and
  reads as "no effect".
• Run check-versions.sh only after git add — it enumerates with git ls-files and cannot
  see an unstaged file, so it passes by not looking.
• Never quote a local build id in a doc. A local build never matches the canonical id anyway
  (hazync#88 — the build path is baked into panic metadata).

What NOT to spend the card on

Faster/bigger cards, cheap tiers, wire compression, NDEBUG, ECMULT_WINDOW, C/C++ LTO, a newer
risc0 — each closed, each with the number that closed it.

And one item that may outrank E9

#139's packer refit — worth 2.36×, host-only, no METHOD_ID, no fidelity cost, and a
prerequisite for #139 rather than a follow-up. That's now built in #190.

Docs only.