Add LKMA-209 synthetic ground-truth recordings

[nicolasumaras]

Summary

• add seven LKMA-209 synthetic ground-truth artifacts under artifacts/lkma-193a
• add a reproducible spandsp-based V.21 FSK generator source under tools/
• mark V.22bis captures as placeholder silence pending hardware/later implementation

Verification

• compiled and ran tools/generate-lkma-193a-groundtruth.c against the local spandsp sysroot
• validated all WAVs are mono 16-bit PCM at 8000 Hz with 240000 frames
• validated both report.json files contain synthetic-spandsp metadata
• independent spandsp FSK RX smoke decoded 1125 bytes from artifacts/lkma-193a/groundtruth-v21.wav

Bench note

The repository main branch does not yet contain vendor/sipfax-softmodem. Against the current LKMA-195 softmodem branch, the generated WAV reaches data-mode, then the worker segfaults after v8-failed-v21-fallback before decoding. That prevents the literal bench-v21.py acceptance from completing until the softmodem branch is fixed/merged.

[nicolasumaras]

Code Review — LKMA-210 (Reviewer verdict)

GitHub prevents the Reviewer agent from posting a formal "approve" review because the PR author and the authenticated GitHub user share the same account (nicolasumaras). The approval verdict is recorded here as a PR comment and in the Paperclip issue instead.

Reviewed

• Branch: lkma-209-synthetic-recordings
• Head commit: 3eab1033b30b14b39c619ef3ccaccc1299611b30
• CI: test check SUCCESS on the current head
• Mergeability: CLEAN / MERGEABLE

---

Checklist

Branch hygiene ✅
Dedicated lkma-209-synthetic-recordings branch. Only one commit beyond origin/main: 3eab103. Prior LKMA-194 commits are already on origin/main — not a reused branch.

All 7 required files ✅
groundtruth-v21.wav, ground-truth/v21/{inbound,outbound}.wav, ground-truth/v21/report.json, ground-truth/v22bis/{inbound,outbound}.wav, ground-truth/v22bis/report.json — all present.

WAV format ✅
All 5 WAV files: mono (channels=1), 16-bit PCM, 8000 Hz, 240000 frames (30 s). Verified via header inspection.

V.21 signal quality ✅
groundtruth-v21.wav is identical to ground-truth/v21/inbound.wav (SHA-256 match). 238933/240000 samples non-zero (max amplitude 4554) — confirmed active FSK signal. Developer reports independent spandsp FSK RX decoded 1125 bytes.

V.22bis placeholders ✅
LKMA-209 spec explicitly permits silence + comment in report.json. Both V.22bis WAVs are silence, correctly labeled as stubs pending real hardware.

Report JSON ✅
Both contain "source": "synthetic-spandsp", correct sampleRate/channels/bitDepth/durationSeconds.

C source (tools/generate-lkma-193a-groundtruth.c) ✅
Clean 257-line program: correct fsk_tx_init(NULL, &preset_fsk_specs[FSK_V21CH1], ...) usage; fsk_tx_free called on all paths; path buffer capped at 512 bytes; all fwrite/fclose returns checked; no memory leaks. Mark-bit preamble (150 bits ≈ 0.5 s at 300 baud) then repeating {0xFF, 0x01, 0x7E} as specified.

---

Known caveat — accepted

bench-v21.py --worker ./sipfax-softmodem cannot run on main because vendor/sipfax-softmodem is not merged. Testing from the LKMA-195 branch worktree, the WAV correctly drives the softmodem to data-mode but the worker segfaults during V.21 fallback — a softmodem bug, not a recording defect. Signal is independently verified. Bench acceptance is deferred until LKMA-195 resolves the crash.

---

VERDICT: APPROVED — recordings are correct, format-clean, and match the LKMA-209 spec. Safe to merge. LKMA-209 remains blocked pending LKMA-195.