perf: reduce Market table resize rerenders

[huhuanming]

What changed

• Keep Market desktop column descriptors stable across the gtLg and gtXl breakpoints, then filter only the active 8/9/11-column set.
• Extract table cell rendering into a memoized MemoTableCell, allowing unchanged cells to skip React and Tamagui style resolution when a responsive column is added.
• Add a production Market resize benchmark with equal-distance controls, visible-column fixture assertions, stale-build detection, thresholds, and aggregation.
• Add Windows support for browser discovery, repository Yarn process spawning, agent checks, and the Web postbuild step.
• Document the investigation, rejected alternatives, benchmark procedure, and follow-up work in the Tamagui performance handoff.

Why

The previous md resize case replaced the desktop/mobile shell, so it was not a fair measure of Market list performance. The corrected benchmark keeps the desktop shell stable and measures the real 8 -> 9 column transition around 1024 px.

At the breakpoint, useColumnsDesktop() previously rebuilt every column descriptor. Each memoized row received a new columns array and reconstructed every existing cell, causing repeated React work and deep Tamagui style resolution.

Impact

Production build, three-run median, 20 measured transitions per scenario:

Metric	Baseline	Candidate	Change
Scripting	3,926 ms	895 ms	-77.2%
Main-thread task duration	5,840 ms	2,534 ms	-56.6%
Long-task total	4,203 ms	644 ms	-84.7%
Maximum frame	300 ms	100 ms	-66.6%
Dropped-frame estimate	265	79	-70.2%
Style recalculation	992 ms	904 ms	-8.9%

The equal-distance non-breakpoint control scripting time also improved by 15.3%. Visible behavior remains 8 columns below 1024 px, 9 columns at gtLg, and 11 columns at gtXl. Hidden columns are not kept mounted.

Validation

• yarn agent:check --profile commit
• 8 focused Jest suites, 103 tests
• Production three-run Market control/cross comparison
• XL 9 -> 11 column stress run
• Market cold-start and initial bundle budgets
• Exact visible column counts, header/row alignment, horizontal overflow, and token-route click-through checks
• git diff --check

Two alternatives that kept all 11 cells mounted using Tamagui responsive props or native CSS media classes were measured, regressed performance, and were reverted.