FP hardware survey — fleet capture campaign (tracking)

[MikeGrier]

FP hardware survey — fleet capture campaign

Coordinates the cross-machine floating-point hardware survey produced by
crates/fp-hw-survey. The tool captures native scalar FP results + exception
flags through per-architecture inline asm (not libcore math), then merges
captures from many machines to surface the rows where hardware actually
disagrees. See crates/fp-hw-survey/README.md → Collecting captures across
the fleet.

Why this campaign

• Confirm intra-ARMv8 determinism. The Arm ARM defines frecpe/frsqrte/
  frecpx/fmulx via exact pseudocode; distinct ARMv8 vendors should produce
  bit-identical estimates. A zero-divergence merge across AArch64 machines is
  the expected, informative result — positive confirmation of conformance.
• Map the genuinely divergent axes. Cross-architecture x86-64 ↔ AArch64
  semantics (fmax/fmin NaN handling, float→int saturation, flush-to-zero
  edges), and any single-precision estimate difference attributable to a
  FEAT_RPRES / FPCR.AH context gap (rpres/afp are recorded in each
  capture header precisely so such a row is read as a feature difference, not an
  erratum).

Target SKUs

• Apple Silicon (M-series, arm64/macOS) — fp-hw-survey capture: Apple Silicon (M-series, arm64/macOS) #166
• Snapdragon X-series (arm64/Windows-on-ARM) — fp-hw-survey capture: Snapdragon X-series (arm64/Windows-on-ARM) #167
• Ampere Altra / AWS Graviton (arm64/Linux) — fp-hw-survey capture: Ampere Altra / AWS Graviton (arm64/Linux) #168
• x86-64 baseline (Intel + AMD) — fp-hw-survey capture: x86-64 baseline (Intel + AMD) #169

Acquisition (manual, per machine)

No CI capture job — the interesting SKUs are physical hardware a hosted runner
does not represent. Each machine owner builds --release, runs info +
selftest (capture refuses to write if the self-test fails), then capture --label <cpu-and-os>, and submits the NDJSON plus the one-line header on the
matching per-SKU issue above.

Ingest (offline, on one machine)

fp-hw-survey merge --out divergences.ndjson capture-*.ndjson

Commit only, never the raw multi-hundred-MB captures:

• divergences.ndjson — the merge output (expected empty for intra-ARMv8
  estimate ops; an empty file is the positive result).
• a provenance table — one row per contributing capture copied from its
  header: label, cpu, os, features, captured_utc, row count,
  tool_version.

Raw captures are retained out-of-band (issue attachments / artifact storage).

Close-out

Paste the merge stderr summary (machine count, aligned-key count, divergence
count, per-op breakdown) into this issue and check off the contributing SKUs
above.