docs(filetests): retire four stale claims found by the f32 roadmap's G3 sweep

[Yona-Appletree]

The f32 native-math roadmap's G3 sweep — M9's roadmap-wide cleanup, reassigned to M8 along with the gate (Yona, 2026-08-02, since M9 merged while M6–M8 were still in flight).

The sweep hunted the failure class M7 P6 found: a statement true when written that quietly went false. Four turned up, all in filetest docs. No disposition changed and no baseline was regenerated.

What was stale

Where	Claimed	Actually
src/targets/mod.rs	wasm.f32 is out of DEFAULT_TARGETS because @lpfn/@glsl imports resolve to Q32 builtin ids	Obsoleted by M5 — resolve_builtin_id_for_mode landed with tests pinning "no import ever resolves across modes", including the load-bearing pointer-parameter case
README.md	Held back "pending review gate G1"; 52 files blocked; _f32 bodies are "stubs that round-trip through Q32::from_f32_wrapping"	G1 passed 2026-07-31; the count is 2; the stubs are gone
README.md	Explicit-only target list	Omitted xtn.f32 / xtlpn.f32, which M8 registered
filetests/uniform/struct.glsl	"Which side is authoritative is a G1 question; @broken until answered"	G1 answered it. float.md §5 classifies both operations — max with a NaN operand, and normalize(vec3(0)) — as Unspecified

The one that matters beyond docs

wasm.f32's exclusion from CI is now inertia, not a blocker. Measured on this base:

                  pass    fail   unimpl  unsupported  compile-fail
      wasm.f32    6318       0        2          199             0
6345/6345 tests passed, 2 expected-failure, 850/850 files passed, 13.34s

Unlike the xtn.* targets it needs no cross-target artifact — wasmtime is already a test dependency. Not flipped here: promoting it to DEFAULT_TARGETS is a CI-cost decision for G3, and roadmap Q6 did want it default. This change only stops the docs from stating a fixed defect as current.

Left deliberately

uniform/struct.glsl's @broken(wasm.f32) should be retired rather than carried — float.md §5 rule 3 says unspecified behavior is never asserted. Doing that correctly needs the behavior of all five f32 targets, and the q32 assertion must survive, so it is recorded as a G3 follow-up rather than done blind.

Validation

cargo check -p lps-filetests                    ✓
just test-filetests → 31572/31572, 850/850 files, unchanged
scripts/filetests.sh --target wasm.f32 → 6345/6345, 850/850 files, 13.34s
just build-fw-esp32c6 && just fw-esp32c6-size-check
  → 2880000 B / 3145728 B — headroom 265728 B (margin 65536 B)

Full G3 artifact (not in this repo): ~/.photomancer/planning/lp2025/2026-07-30-1745-f32-native-math/g3-roadmap-final-review.md

🤖 Generated with Claude Code