perf(ice): zi ice tokenizing is 20% slower since #527

[ss-o]

Problem

An A/B on 2026-09-18 (fresh zsh -f per sample, alternating order, 3 warmups, 30 samples) between main 56715fd (v2.1.0) and next b4ab453 shows one case outside noise: 200 zi ice calls carrying 13 ices each take 151 ms on main and 182 ms on next, a +20.5% median regression, while every other case (source, light and tracked load, turbo queue plus burst, unload) is within ±1%. The change is #527: .zi-ice now consults ZI[nval-ice-list] and inspects the remainder for every matched word.

Absolute cost is about 0.15 ms per zi ice call, roughly +8 ms on a startup file with 50 ice calls, so this does not block the promotion; it is the kind of drift the benchmark suite (#553) exists to catch.

Expected behavior

The common case (a valued ice, or a bare no-value ice) stays on the previous path; the extra check runs only when a matched no-value ice name has a non-empty remainder. Measure per-word cost with the benchmark's ice case before and after.

[ss-o]

Project 28 triage pass. Blank values filled on this issue (existing values left unchanged):

• Item Type: Performance. A +20.5% median timing regression in .zi-ice.
• Impact: Low. The issue measures about 0.15 ms per zi ice call, roughly +8 ms for 50 calls.
• Effort: S. A fast-path condition in .zi-ice plus a before-and-after measurement.
• Priority: P3 - Low. The issue says it does not block the promotion.
• Workstream: Human delivery. Ordinary performance work.

[ss-o]

Project 28 triage pass (2026-10-02): stays in Triage.

• Premise check on next at 3fa8433 (premise not established): The fix(ice): stop tokenizing at a no-value ice name followed by text #527 check is unchanged at zi.zsh:2605@3fa8433 ([[ ${match[2]} = (${~ZI[nval-ice-list]}|svn) && -n ${match[3]#(:|=)} ]] && break runs for every matched word; no fast path). Measurement not reproduced: main@4f2c0a7 already carries fix(ice): stop tokenizing at a no-value ice name followed by text #527 (promoted by chore: promote next to main #567), so a 7-round alternating A/B of 200 zi ice calls x 13 ices (main 174-218 ms, next 173-195 ms) shows no delta as expected; a v2.1.0 clone for the real baseline was denied by the runtime.
• Proposed disposition: implement, once the open points below are agreed.
• Unknowns and next investigation: Re-measure against v2.1.0 (56715fd) with benchmarks/run.zsh ice case now that the suite exists; confirm the +20% still holds, then add the fast-path condition and report before/after with the same case.