v0.0.6

Other

• RFC 4587 §6.2.1 offer/answer negotiation helper
• fix truncated 1? TCOEFF prefix panic (daily-fuzz finding)
• criterion suite for transform / encode / decode hot paths
• scrub decorative external-implementation attribution
• second cargo-fuzz target for RTCP compound parser
• cargo-fuzz decoder harness + daily workflow

Added

• SDP offer/answer negotiation helper (RFC 4587 §6.2.1). The
  sdp module gains the free function negotiate_answer(offer, our_capability) -> Result<H261FmtpParams, SdpError> that computes
  the §6.2.1 answer parameters from a received offer and our
  local capability:

  • Picture-size intersection. Only sizes both peers advertise
    survive into the answer; a disjoint pair (e.g. CIF-only offer vs
    QCIF-only capability) errors with SdpError::NoPictureSize,
    matching §6.2.1's "SHALL specify at least one supported picture
    size".
  • MPI per shared size. §6.1.1's MPI is the minimum picture
    interval, so 29.97 / MPI is the upper bound on frame rate.
    The answer carries MPI = max(offer.MPI, our.MPI) per shared
    size, i.e. the more restrictive bound binds.
  • Annex D (D). §6.2.1: "This option MUST NOT appear unless
    the sender of this SDP message is able to decode this option."
    The answer's D=1 requires both offer.d == Some(true) AND
    our_capability.d == Some(true); otherwise D is omitted from
    the answer (matching §6.1.1's "SHOULD NOT be used … if not
    supported").
  • RFC 2032 fallback. §6.2.1: "If the receiver does not specify
    the picture size/MPI parameter … assume that such a receiver is
    able to support reception of QCIF resolution with MPI=1." The
    helper applies that fallback automatically (equivalent to
    H261FmtpParams::rfc2032_fallback()) when the offer carries no
    picture-size parameter. The fallback is not applied to
    our_capability — that side is local and should be supplied
    explicitly.

  The companion method H261FmtpParams::preferred_picture_size()
  returns the preferred receiver mode per §6.2.1 ("Parameters offered
  first are the most preferred") — Some(SourceFormat::Cif) when CIF
  is advertised (matching format_value's CIF-before-QCIF emission
  order from the §6.2.1 worked example), Some(SourceFormat::Qcif)
  when only QCIF is, else None. Eight new tests cover the
  intersection / MPI-max / disjoint-sizes / Annex-D / RFC-2032-
  fallback / format round-trip / max-frame-rate / validate-passes
  paths; the negotiation example also runs as a doctest.

Fixed

• Decoder panic on a truncated 1? TCOEFF prefix (round 175,
  surfaced by the scheduled daily decode_h261 fuzz harness).
  decode_tcoeff(.., is_first = false) saw a bit-reader where exactly
  one bit remained and that bit was 1. The function took the
  b0 == 1 branch and then peeked two bits from
  peek >> (avail - 2), where avail = 1 caused an unsigned
  underflow → attempt to subtract with overflow panic under debug
  / ASAN builds. The two-bit peek is now gated behind avail >= 2
  and the call returns Error::invalid("h261 tcoeff: truncated 1? prefix") on the malformed input, restoring the public-surface
  contract from the fuzz harness: every call returns — no panic, no
  abort, no out-of-bounds. New regression test
  tcoeff_truncated_one_bit_does_not_panic covers it on stable Rust.

Added

• Criterion benchmark suite (benches/transform, benches/encode,
  benches/decode). Round 175 (depth-mode) wires up criterion = "0.5"
  as a dev-dependency and registers three harness = false bench
  binaries so future optimisation rounds have a recorded baseline to
  A/B against:

  • transform times the 8×8 inverse / forward DCT block hot path —
    fdct_intra + fdct_signed (encoder forward pass) and
    idct_intra + idct_signed (decoder inverse pass). One block per
    iteration; throughput reported in samples so per-sample cycle-
    equivalents land naturally.
  • encode times whole-picture encode through the production
    encode_intra_picture / H261Encoder::encode_frame paths in four
    scenarios: QCIF intra-only (no ME), single-P from a pre-built I
    reference, I + 3 P chain (full rate-controller carryover), and CIF
    intra (the 4× area test).
  • decode times whole-picture decode through H261Decoder::send_packet
    • receive_frame, mirroring the encode scenarios. Each decode
      bench runs the in-crate encoder once during setup to produce a
      real elementary stream, so the timed loop measures the decoder
      alone.
      Every benchmark synthesises its YUV source inline from a
      deterministic striped pattern plus low-amplitude xorshift noise —
      no on-disk fixtures, no third-party CLI, no docs/ files read at
      bench time. cargo bench -p oxideav-h261 --no-run doubles as a
      compile-only CI regression guard via the existing matrix.

• Second cargo-fuzz target — RTCP compound parser. New
  parse_rtcp_compound fuzz target drives arbitrary fuzz-supplied bytes
  through the public RTCP parser surface (parse_compound,
  parse_report, parse_sdes, parse_bye, parse_app) so the §6.1
  compound walk (16-bit-length advance), the SR/RR fixed header + RC
  block walk, SDES chunk + item walk (including the PRIV inner 8-bit
  length), BYE reason-string length-prefix, and APP name/data 32-bit
  alignment are all exercised against bytes whose shape the fuzzer
  dictates. Same contract as the existing decode_h261 target: every
  call must return — no panic, no abort, no integer overflow (in debug
  / ASAN builds), no out-of-bounds index, no allocator OOM. The seed
  corpus under fuzz/corpus/parse_rtcp_compound/ contains nine valid
  datagrams (empty RR, SR with no blocks, SR with one block, RR with
  two blocks, SDES CNAME, BYE with reason, APP with PING payload, and
  two compound packets). tests/fuzz_seed_corpus_rtcp.rs drives the
  same logic on stable Rust against the corpus plus several adversarial
  in-line buffers (lying header length, zero-length advance, truncated
  compound, SDES PRIV length overflow, BYE reason overflow, APP at the
  5-bit subtype maximum, unknown PT=205) so a regression in the public
  parser surface trips an existing CI lane rather than waiting for the
  daily fuzz run.