Grayscale monochrome AVIF BLE photo pipeline with text-region crop

[nic-olo]

Scope

BLE photos exist to be transferred to the phone for text recognition, not viewing. This reworks the glasses-side compressAndSendViaBle() pipeline around that:

1. Grayscale luma pipeline (GrayscaleBleProcessor) — subsampled RGB_565 JPEG decode, row-wise luma extraction, contrast stretch, single-channel unsharp, bilinear resize. Every intermediate buffer is 1 byte/pixel instead of the old full-res ARGB decode → createScaledBitmap → RGB sharpen chain. Peak transient memory per BLE photo drops from ~36 MB to ~4 MB.
2. Monochrome (YUV400) AVIF encode from the raw luma plane — new encodeAvifMono() in a Mentra fork of awxkee/avif-coder 1.7.3 (additive-only patch; same package/classes, drop-in AAR in asg_client/app/libs/, patch + rebuild instructions in asg_client/third_party/avif-coder-mono/). No Bitmap/RGBA round-trip anywhere; the bitstream carries no chroma planes. Also exposes AOM speed (cpu-used): BLE_AVIF_ENCODE_SPEED = 8 benchmarked ~2.6x faster than the plugin default on the software AV1 encode that dominates BLE photo latency (4–5.5 s of ~7.4 s end-to-end on MTK8766).
3. Text-region ROI crop (TextRegionDetector) — Sobel edge-density grid on a ~320 px decode; crops to the concentrated text region (10% pad) so the byte budget concentrates on what OCR reads. Conservative: weak/spread signal or a >80%-area crop falls back to full frame; any detection failure degrades to previous behavior.

Kill switches: ENABLE_GRAYSCALE_BLE_PHOTOS (falls back to the legacy full-color YUV420 path, kept verbatim) and ENABLE_BLE_TEXT_ROI_CROP. Size logging now prints original and AVIF bytes/KB per photo.

No SDK API, transfer-protocol, WiFi-path, or gallery changes. Phone-side code is unchanged except a new test fixture.

Test evidence

• Encoder-stack probe with the exact bundled versions (libheif 1.17.6 + AOM 3.8): monochrome encode round-trips (colorspace=monochrome, no chroma planes); speed=8 encoded 1920x1440 in 75 ms vs 198 ms host-side, with comparable size.
• Independent decode check: avifdec (libavif + dav1d, the same decoder family the phone's HeifCoder bundles) decodes the mono output as Format: YUV400 into a correct 8-bit grayscale PNG.
• scripts/check-android-compile.sh asg and full :app:testDebugUnitTest green, including new Robolectric suites for GrayscaleBleProcessor (aspect fit, contrast, ROI crop, gray rehydration) and TextRegionDetector (concentrated block detected, flat/noise frames rejected, source-coordinate mapping).
• scripts/check-android-compile.sh bluetooth-sdk green; phone-side AvifExifStripperTest extended with a monochrome+EXIF fixture generated by the exact glasses encoder stack — strip + JPEG conversion pass.

Outstanding hardware verification

Requires physical devices (not possible from this environment):

• Glasses: real encode time/size at each tier (⏱️ Compression took / ✅ Compressed photo for BLE logs), ROI crop behavior on real documents/scenes.
• iOS: UIImage(data:) render of a monochrome AVIF on iOS 16/17/18 devices (Android decode is covered by the dav1d-family check above). If a specific iOS version misbehaves, flip ENABLE_GRAYSCALE_BLE_PHOTOS or ship YUV420 gray via the legacy path.
• End-to-end OCR sanity check on a photographed document.

  

[github-actions]

📋 PR Review Helper

📱 Mobile App Build

✅ Ready to test! (commit ec8aac8)

📥 Download APK

🕶️ ASG Client Build

✅ Ready to test! (commit ec8aac8)

📥 Download ASG APK

---

🔀 Test Locally

gh pr checkout 3413

[github-actions]

PR Agent Orchestrator State

{
  "cycle": 3,
  "fixRound": 0,
  "totalReviewerRuns": 3,
  "consecutiveNoNewReviews": 0,
  "openFindings": [],
  "resolvedFindings": [],
  "nitFindings": [
    {
      "id": "bb5f98",
      "fingerprint": "asg_client/app/libs/avif-coder-mono-1.7.3-mono1.aar:bcada628f89e",
      "source": "standards",
      "severity": "nit",
      "file": "asg_client/app/libs/avif-coder-mono-1.7.3-mono1.aar",
      "line": 0,
      "message": "~8.36MB prebuilt AAR committed directly to git (not Git LFS) permanently bloats repo history. Well-documented in third_party/avif-coder-mono/, but consider Git LFS or a published JitPack fork as the README itself suggests. Non-blocking.",
      "status": "open",
      "introducedCycle": 2,
      "lastSeenCycle": 2
    }
  ],
  "phase": "discovery",
  "status": "in_progress",
  "lastPair": [
    "bugbot",
    "standards"
  ],
  "stagnationFixRounds": 0,
  "lastOpenCount": 0,
  "fingerprintReopenCounts": {},
  "mutedFingerprints": []
}

[github-actions]

🤖 PR Agent Review — cycle 3

✅ No blocking findings · 0 blocking · 1 nit
Reviewers this cycle: bugbot, standards

Claude — standards — ✅ approve

Commander, recon complete. The mission is resolved — I've inspected the governing AGENTS.md files (root + asg_client/), traced the changed source across files, and confirmed runtime wiring. Here's the after-action report.

Review — PR #3413 (BLE monochrome AVIF photo pipeline)

Standards & conventions

• Java naming/style (asg_client): Compliant. New classes GrayscaleBleProcessor, TextRegionDetector are PascalCase; methods camelCase; constants (CENTER_W, ANALYSIS_MAX_DIM, ENABLE_GRAYSCALE_BLE_PHOTOS, BLE_AVIF_ENCODE_SPEED) UPPER_SNAKE_CASE; 4-space indent per asg_client/AGENTS.md. Data-holder fields (LumaImage.luma/width/height) aren't m-prefixed, but that matches the pre-existing BleParams holder in the same file — consistent, not a regression.
• Javadoc: Public/package methods and classes are documented. Good.
• Wiring verified: encodeAvifForBleMono resolves HeifCoder, AvifBmffExifInjector, buildExifApp1Segment, hasImuMetadata (all present); encodeAvifMono(luma,w,h,stride=w,q,speed) matches the patched Kotlin signature and buffer-size precondition (w*h >= w*(h-1)+w) holds. Locale/File imports present for the new size logging. Legacy color path preserved behind ENABLE_GRAYSCALE_BLE_PHOTOS kill switch.

Testing (per AGENTS.md — BLE features need unit coverage + E2E path)

• New logic is covered: GrayscaleBleProcessorTest (aspect fit, contrast, ROI crop, gray rehydration) and TextRegionDetectorTest (concentrated block, flat frame, whole-frame noise, source-pixel mapping), plus a phone-side mono-AVIF decode fixture test in AvifExifStripperTest. Adequate.

Commit / PR hygiene

• No AI co-author trailers and no "Generated with" lines. The Co-authored-by: Nicolo Micheletti trailer names a human — allowed. Commit subjects are imperative and scope is focused on the mono AVIF pipeline. Compliant.

Security

• No committed secrets. No cloud/Docker/Mongo changes, so the localhost-binding rule is untouched.

Nit (non-blocking)

• asg_client/app/libs/avif-coder-mono-1.7.3-mono1.aar is an ~8.36 MB prebuilt binary committed directly to git history (not Git LFS; check-attr shows no LFS filter). It's well-justified and documented (third_party/avif-coder-mono/ README + patch + rebuild steps), and there's no repo LFS convention to violate — but a multi-MB blob permanently bloats history. Consider LFS or a published JitPack fork (the README already flags this as the longer-term plan).

Verdict: no blocking issues or regressions — the diff is self-contained, tested, and standards-compliant. Approving.

_{Updated automatically by the PR Agent Orchestrator each review cycle. Nits do not block merge.}

[github-actions]

bugbot run