fix(graphics): retain Progressive difference tiles

[Marc-André Moreau (mamoreau-devolutions)]

Retain quantized DWT coefficients per Progressive tile so difference updates compose with their matching surface reference while progressive codec-context state remains isolated.

Reject difference tiles that lack a retained reference instead of decoding them against zeros.

Keep retained surface references across codec grid replacement and ResetGraphics; only deleting the surface releases them.

[Copilot]

Pull request overview

Adds Progressive RemoteFX difference-tile decoding with retained per-context tile references.

Changes:

• Defines the difference-tile flag.
• Accumulates difference coefficients and rejects missing references.
• Adds difference-tile regression tests.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
crates/ironrdp-pdu/src/codecs/rfx/progressive.rs	Defines and tests the difference flag.
crates/ironrdp-graphics/src/progressive.rs	Retains references and decodes difference tiles.

[github-actions]

Adds decode support for RFX_TILE_DIFFERENCE in Progressive TILE_SIMPLE/TILE_FIRST tiles: retains a tile's quantized DWT coefficients as a reference, sums a difference tile's decoded delta onto it, and returns MissingTileReference instead of silently decoding against zero when no prior tile exists. Well-scoped, keeps the existing decode_first signature, and adds two targeted tests (cross-context accumulation, missing-reference error). Arithmetic and error logic match the supplied protocol handoff's conformance analysis for MS-RDPEGFX 3.2.8.1.4 and 3.3.8.2.1/.1. Two secondary handoff points (context- vs surface-scoped reference lifetime, saturating vs wrapping addition) are real but low-impact. One added point: a new variant on a non-exhaustive-unmarked public error enum is a legitimate but expected semver note.

Protocol analysis: accepted — Re-derived the mappings independently: TILE_FLAG_DIFFERENCE=0x01 matches the documented flags bit for TILE_SIMPLE/TILE_FIRST; the zero-then-add-back-reference sequence is algebraically equivalent to DecDwtQ = Ref + DecProgQ*PQF (difference) vs DecProgQ*PQF (original), with sign state always cleared; gating on pass != 0 (else MissingTileReference) defensibly enforces the encoder-side 'first send is always original' rule. The two potential_discrepancies are correctly low-confidence: context-vs-surface keying predates this diff, and saturating_add is a safe divergence from unspecified overflow behavior, not a wire defect.

[Yevhen Salitrynskyi (ysalitrynskyi)]

Same bug as #1701. Keying the reference by surface and keeping it across ResetGraphics is the right model (MS-RDPEGFX 3.3.1.3) — mine conflated it with the codec context.

One substantive point: rejecting a difference tile that has no retained reference. 3.3.8.2.1 makes the zeroing conditional — "if the tile is an original tile (not a difference tile), then the tile MUST be zeroed out in the current frame" — but the add unconditional: "The result of the entropy decode operation MUST be added to the current frame." No error case is defined, and FreeRDP decodes against a zero-initialised current. The error also propagates out of decode_bitmap, so one such tile discards every other tile in the same PDU and handle_wire_to_surface2 turns it into a channel error. Decoding against zero degrades one tile instead.

Nit: FLAG_TILE_DIFFERENCE would match the existing FLAG_DWT_REDUCE_EXTRAPOLATE.

Tests in #1701 you may want to lift: TILE_FIRST end to end (only TILE_SIMPLE is exercised here), the reduce-extrapolate band layout, and a difference tile arriving after an upgrade pass.