fix(decoder): byte-exact sub-range (col_range) IDWT at component boundaries

[osamu620]

Summary

The reuse-path sub-range horizontal IDWT (the column-range / JPIP-zoom decode) was not byte-exact with full-width decode when a window edge fell on a component boundary — which surfaced as the scalar build's cr_ht_06_reuse conformance failure.

idwt_1d_row_inplace_range clamped its lifting-pass range to [0, width-1]. At a component boundary (col_lo == u0 or col_hi == u1) that dropped the PSE-region writes the full-width kernel makes, leaving the first/last ~2-3 output columns wrong; the per-level error then compounded down the streaming-IDWT cascade. Subsampled chroma was worst hit (its window clamps to the component edge at every level). Interior window edges, which keep a finite real-neighbour margin, were already exact.

Fix: at a component boundary, run the lifting pass to its natural extent (writing into the row's PSE prefix/suffix slack exactly as the full-width kernel does) instead of clamping; keep the finite margin only for interior edges. The change is in the shared scalar/interleave fallback that every ISA build uses for coarse DWT levels (N < 16), so it covers all ISAs.

Verification

• 449/449 conformance on AVX2 and scalar (scalar previously failed cr_ht_06_reuse).
• Full-image decode is byte-identical (the full-width fast path is untouched; the change only affects an actual sub-range window).

Deferred follow-ups

Two related improvements were prototyped and intentionally left out of this PR. They exposed a separate, pre-existing latent fault in the windowed-decode path that reproduces only on certain CI toolchains (GCC 13 / MSVC / AppleClang, non-deterministically) and not under local AddressSanitizer (pooled allocator disabled, so all buffers redzoned — decoder is memory-clean) or valgrind, so they need dedicated investigation:

• Wiring set_col_range to actually apply on the non-reuse line-based path (it is currently a silent no-op there, because it runs before init_line_decode builds the tree).
• Broader edge / narrow-window coverage in col_range_compare's reuse probe.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WaH1Ay9hPfnqfHsJ7wxh8M

[Copilot]

Pull request overview

This PR fixes a correctness mismatch in the decoder’s windowed (column-range / JPIP zoom) horizontal IDWT on the reuse path when the requested window edge lands at/near a component boundary, ensuring byte-exact output versus full-width decode and preventing error compounding in the streaming IDWT cascade.

Changes:

• Adjust sub-range horizontal IDWT range selection so that when a window edge is within filter support of a component boundary, the lifting pass runs to its natural extent (matching full-width PSE-region writes).
• Preserve the finite-neighbor “widen” margin behavior for interior window edges.
• Add <cstdint> include to support INT32_MIN/INT32_MAX sentinels used by the range-clip logic.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.