perf(decode): batched refill + multi-literal decode inside the fastloop (revisit #2630/#2669)

[kim-em]

Revisit batched bit-refill and multi-literal-per-refill decoding inside the fastloop, now that the two prior negative results (#2630, #2669) have had their premises removed. Strictly sequenced after the packed length/distance entries and the fastloop split.

Why the prior negatives no longer apply

Both earlier attempts measured the right change on the wrong shape, and both post-mortems said so explicitly:

• perf(explore, needs lean4#14053): decode wide bit-refill — ugetUInt64LE in refill #2630 (wide refill) — measured neutral, because the eager loop tops up ~1 byte at a time (k≈1), so a single 8-byte load replaces a single byte load. Its own conclusion: "a batched-refill restructuring is the only thing that could change that."
• perf(decode): decode multiple literals per refill (Phase 2 of #2668) — deferred, depends on #2670 #2669 (multi-literal) — regressed on every variant, because pre-packed-entry the "continue the batch?" test cost a full multi-read + multi-compare guard on every literal, wasted whenever the next symbol was a match (~25-30%). Its own conclusion: depends on a packed entry making that test a single bit branch.

After the fastloop split removes the per-refill input bounds check, and the packed length/distance entries make the fast-literal test a single bit test, the premises flip:

• refill once per iteration with one 8-byte wide load (k≈7 now, not 1), then
• decode 2-3 literals, or a literal + a full match, from the ~56 buffered bits — libdeflate does up to 3 literals per refill with next-entry preloading; miniz does 1 refill per 2 literals.

Method

Keep the same-worktree A/B discipline from #2630 (a cross-worktree baseline once produced a 37-point phantom). Spike with the equivalence sorry'd. If, even on the restructured loop, our codegen still measures this neutral, close it with the measurement — the fastloop-split gains are retained regardless.

Dependencies

Hard: requires the packed length/distance entries (single-bit literal test) and the fastloop/slowloop margin split (no per-refill input check) to be in place first.

🤖 Prepared with Claude Code

[kim-em]

Depends on #2798 (packed length/distance entries → single-bit fast-literal test) and #2799 (fastloop split → no per-refill input check). Do not start until both have landed and been measured; the whole point is that those two remove the premises that sank #2630 and #2669.