Performance Researched

Researched mechanisms

Reviewed: 2026-11-01 · Source: b7140848 · Availability: current main; stable/main boundaries are stated below.

These experiments have evidence worth preserving but are not delivered as the described end-to-end mechanism in reviewed main. GitHub state and recommendation differ: HOLD denotes incomplete/inconclusive qualification, while SKIPPED can record a stopped route. Neither alone means a feature has no possible value.

Research	Disposition / lesson
#83 descriptor pool, #88 state reuse	Closed unmerged; removing a named finish can displace waits rather than improve total progress.
#87 direct vertex writes	Closed unmerged; part of its direction is incorporated in delivered #89. Current vertex route.
#193 attenuation lookup	Measured4096-entry candidate closed unmerged: cadence+0.70%, CPU−0.62%, p99+6.32% worse. Detail.
#195 stereo nibble table	Closed unmerged; separate from delivered decoded-block cache #187.
#184/#194/#196 prepared controls	Closed unmerged; narrow callback-local CFG_FMT #191 is delivered, broader persistent cache is not.
#330 cubemap copy, #339 overlap, #340 empty publication clocks	Stopped/closed candidates; retained useful work counters did not establish acceptable whole-game outcomes. Preserve adverse tails and incomplete qualification.
#318 occlusion	Closed unmerged; directional+2.305% Deck cadence pair has mixed tails/worse max, not zero potential or a qualified broad gain.
#319 GL recycling	Open/HOLD; original campaign collected, B3/B4 stale traces need diagnosis before qualification.
#55 interrupt/STI shadow, #92 snapshot reproduction	Measured research; high frequency is not removable cost; startup load did not reproduce live restore.

#85 freshness and #89 bounded vertex versions are merged/present, so their current home is Implemented, while original PR lessons remain historical. Current research and upstream actions gives partial, removed and unmerged states. PR lessons retains original reasoning.