parityCount: Math.ceil → Math.floor
- Small files (G ≤ 33) now truly get 0 parity generations
- Large files (G ≥ 34) still get proportional parity (≈3%)
- Previously Math.ceil always gave ≥1 parity for any G ≥ 1
neededPackets: K * sourceGenerations → K * totalGenerations
- Aligns with the round-robin scheduler: symbols are interleaved
across ALL generations, so practical minimum is K × totalGens
- Fixes the mismatch where 'needed' showed 32 but actual
decode required ~48 frames (for 2 source + 1 parity case)
Tests: Updated assembly test comments and totalGenerations params
to match new parityCount behavior.
Closes the long-standing 'needed frames count is off because of
outer error correction' issue.
- framesWithQR now counts ALL frames with a valid QR (including duplicates),
not just deduplicated ones. Moved increment before dedup check.
- acceptedPackets already correctly counts linearly-independent packets
(rank increase), capped at K per generation by the decoder.
- neededPackets = K * sourceGenerations, computed from constants.
- Keep generation count (solvedGens/sourceGens) as secondary indicator.
This gives an honest rough progress bar: useful can approach or briefly
touch required while some generations are still partially filled, but the
generation count is the definitive completion signal.
The compact indicator previously showed framesWithQR/acceptedPackets/neededPackets.
Because acceptedPackets counts parity-generation packets and uneven distribution
across generations, it could exceed neededPackets while the file was still not
decoded. This is confusing.
Change the primary compact indicator to generation counts:
framesWithQR / solvedGens / sourceGens gens
This is always meaningful: solved never exceeds required, and when they match,
the transfer is complete. Packet count is kept as a raw secondary number.
- Add outer Reed-Solomon error correction across generations
(3% overhead, 1 parity gen per ~33 source gens)
- Fix progress display: show generations solved/needed instead of
raw packet counts, making outer EC benefit visible
- Add decode timer: tracks elapsed time from first frame to completion
and shows throughput in KB/s
- Add benchmark test confirming outer EC reduces tail waiting time
under biased frame loss
- Remove per-generation hash verification (superseded by outer RS)
- Update deployment base path to /hermes-web-demos/qr/
- Removed session tracking entirely; single global decode state
- Receiver UI: no sessions table, no big progress bar
- Inline stats above video: scanned / useful / need · gen X of Y
- Sender passes filename and mimeType to encode worker
- Packetizer wraps file payloads with [filenameLen][name][mimeLen][mime] header
- Decode worker parses metadata and uses it for download name/type
- Fixed generation counter bug: final progress was skipped when reconstruction
succeeded; now reportProgress is called right before early return
- Frame count now shows: scanned / useful (linearly independent) / minimal needed
- Camera zoom via getUserMedia track.applyConstraints when supported
- Software crop fallback with visible dashed-square overlay on video
- Corner markers on scan region for visual feedback
- qr_decode accepts inversionAttempts option for faster GIF decoding
- Preserved attemptBoth for camera scanning (handles glare/reflections)
- Remove CBOR manifest and manifest.ts entirely
- Fixed 18-byte packet header: sessionId, generationIndex, symbolIndex,
packetType, totalGenerations, dataLength, flags, reserved
- Single hardcoded profile: QR V10, ECC M, K=16, R=8, payload=191 bytes
- Add isText flag (bit 0 of flags) for text vs binary routing
- Remove hash computation and profile selection UI
- Update all tests to match new protocol (33 tests passing)
- Fix generateCoefficients arg order bug in complete.test.ts
- Fix createQRGif dimension bug in tests
- Make prod_roundtrip frame loss deterministic
All 33 tests pass. Build succeeds.
The dedup key was `::`, which
caused manifest fragment 0 to collide with systematic symbol 0. The RLNC
decoder then never reached full rank, so generations never solved and the
receiver appeared to hang.
- Add packetType to dedup key so manifest/systematic/coded symbols are unique
- Add completed session flag to ignore late frames after reconstruction
- Prevent trailing progress message from overwriting complete status
- Add frame_decode and prod_roundtrip tests
- Fix stale closure bug in camera capture loop (scanning state
captured as false, preventing any frame capture)
- Add GIF File input mode to receiver — upload a QR-over-GIF
directly to test the decode pipeline without camera
- Update decode worker to accept imageData or frameData
- Receiver now shows two input modes: Camera and GIF File
Hermes Agent