Packed-average cursor recognizers — 2026-09-22

Pass-5 §P5-3 names loop_try_avg_shift_cursor and loop_try_avg_round_cursor as the two largest near-duplicate recognizers. Both now delegate to loop_try_avg_cursor, with an explicit rounded flag. This removes 109 lines from 07b-loop-match.wat without merging runtime instruction handlers or changing the opt-in generic-copy emission gate.

The shared implementation validates the two source loads, equal masks, shift/add sequence, destination store, three four-byte cursor increments, countdown and branch. Rounded mode alone validates and emits the three pre-shift correction operations and the post-average correction add. It requires seven distinct register roles instead of six. The operation count remains 17 versus 13, and the fused descriptor retains mode 2 versus 1. Each form still emits its complete ordinary final iteration to preserve scratch-register and flag results. New mode branches run during decoding, not for each averaged pixel; no speedup or neutrality claim is made.

Verification:

• test/test-packed-avg-run.js compares fused and ordinary output bytes, all general registers and CF/ZF/SF/OF for indexed, shift-cursor and rounded-cursor forms. It also covers indexed overlap and budget re-entry.
• Five added cursor near misses check unequal masks in both forms, a wrong cursor increment, a wrong correction source and a wrong correction add. Each runs with emission disabled and enabled: neither recognizes, and output/register/flag state agrees. The original indexed near miss remains.
• Both suites pass, including the neighboring test-copy32-counted-run.js checks for overlap, sparse-page crossings, short rows and slice resumption. Fragment balance (107), logical operands, handler stack cleanup, test-tier discovery and whitespace checks pass. The exact-duplicate ratchet remains 138 groups / 531 members; this change removes near duplication, not an exact-duplicate group.

This closes this named near-duplicate pair, not all review recommendations or the broader safety/performance case for enabling copy super-operations. No game benchmark, browser capture or full-build result is implied.