fix(verify): repoint wrapper to archive/v1 paths (#559) + cross-platform FFT probe (#560)

[ruvnet]

Closes #559. Adds empirical root-cause evidence for #560 (left open pending maintainer decision on quantize-before-hash).

#559 — ./verify wrapper points at removed v1/ paths

./verify failed before reaching the proof script on a fresh clone because the wrapper still pointed at the pre-archive layout:

FAIL: Reference signal not found at .../RuView/v1/data/proof/sample_csi_data.json
FAIL: verify.py not found at .../RuView/v1/data/proof/verify.py

Repointed PROOF_DIR, V1_SRC, and two diagnostic strings to archive/v1/.... Three-line core diff matches exactly what @Fewmanism proposed in #559.

#560 — Empirical root-cause probe

Added scripts/probe-fft-platform.py which runs verify.py's hash-relevant scipy.fft.fft + scipy.signal.windows.hamming calls in isolation on a deterministic input (no pydantic Settings stack), so the source of divergence can be located cleanly.

Tested via Tailscale on three real machines:

Platform	numpy/scipy	first 4 PSD floats	sha256
Windows (Intel AVX-512)	2.4.2 / 1.17.1	..., 94.40426770856882, ..., 51.677496924642476	78b3fb…
ruvultra (Linux x86_64)	1.26.4 / 1.14.1	..., 94.40426770856882, ..., 51.677496924642476	41dc56…
ruv-mac-mini (Apple Silicon arm64/NEON)	2.4.4 / 1.17.1	..., 94.4042677085688, ..., 51.67749692464246	9b5e19…

• Win + Linux agree on first PSD/doppler values but produce different SHA-256s → divergence is in later FFT bins, where scipy version's pocketfft SIMD path differs.
• Mac arm64 differs from x86_64 at the first bins at ULP precision (~2e-14 on value ~94 = ~1 ULP).

The divergence is at the SIMD level: NEON (Apple Silicon) vs AVX2/AVX-512 (x86_64) reorder vectorized FP operations differently. IEEE 754 guarantees per-operation determinism, not associativity under reordering — so the docstring at archive/v1/data/proof/verify.py:172 ("platform-independent for IEEE 754 compliant systems") is incorrect.

What this PR does not fix

The architectural fix for #560 is quantize-before-hash in features_to_bytes() (round float64 to a fixed precision well above SIMD-ULP, e.g. 1e-9, before packing) — then regenerate expected_features.sha256 on a canonical CI platform. That changes a published trust-anchor artifact and merits a maintainer call. This PR ships only the path fix + diagnostic probe so #559 unblocks immediately while #560 gets reviewed.

Verification

# Before: FAIL before reaching pipeline (v1/... not found)
./verify
# After:  reaches verify.py and runs the pipeline (path layer healed)

# Cross-platform divergence reproducible:
python3 scripts/probe-fft-platform.py

Test plan

• ./verify on macOS / Linux / Windows: wrapper reaches verify.py and the path-level errors are gone
• scripts/probe-fft-platform.py runs cleanly on Windows + ruvultra Linux + ruv-mac-mini Mac arm64
• Probe output empirically demonstrates the platform divergence claim from Proof replay hash mismatches with pinned requirements on macOS arm64 #560
• Maintainer decision on Proof replay hash mismatches with pinned requirements on macOS arm64 #560 fix shape (quantize-before-hash vs platform-pin vs multi-hash)

Thanks to @Fewmanism for the clean repro + the suggested diff in #559.

🤖 Generated with claude-flow

[ruvnet]

Closing in favor of two cleaner follow-ups now that this session has merged the verify path fix and extracted the probe script:

• ./verify still points at removed v1/ proof paths #559 portion — the wrapper path fix to archive/v1/ shipped via fix: bug triage for #559, #561, #588 #590 (merged earlier today). After PR chore(ci): add fix-markers for recent merges (#559, #561, #588, #593, #590-CI) #606 (fix-marker for ./verify still points at removed v1/ proof paths #559) lands, a regression of the wrapper paths will fail CI.
• Proof replay hash mismatches with pinned requirements on macOS arm64 #560 portion — the probe script (scripts/probe-fft-platform.py) is extracted unchanged into chore(scripts): probe-fft-platform.py — root-cause aid for #560 #607, with the empirical Win + Linux + Mac arm64 evidence preserved in the PR description. The architectural quantize-before-hash decision for Proof replay hash mismatches with pinned requirements on macOS arm64 #560 is left as a separate maintainer call, as originally noted in this PR's body.

Rebasing this PR would leave the verify portion as a no-op and just the probe — but the cleaner audit trail is to land the probe in #607 and close this. Apologies for the churn — #559 was reported, fixed, and shipped in parallel windows.