coli: measured Windows launcher defaults — OMP tuning parity + DIRECT/PIPE/PILOT_REAL (opt-out)

[KingIcyCreamProjects]

On Windows a stock run leaves large measured wins on the table: the glm.c self-exec OMP tuning is Linux/FreeBSD-only, and the PREFETCH/readahead path is a posix_fadvise no-op under _WIN32. This teaches the launcher to setdefault, on win32 only:

• the OMP block (OMP_WAIT_POLICY=active, GOMP_SPINCOUNT, OMP_DYNAMIC=FALSE, OMP_NUM_THREADS=<physical>), same scope + presence-based kill-switch (COLI_NO_OMP_TUNE) as the engine's own tuner; OMP_PROC_BIND/OMP_PLACES are deliberately omitted (and no longer emitted by --auto-tier on win32) because MinGW libgomp rejects affinity outright;
• DIRECT=1, PIPE=1, PILOT_REAL=1 — all lossless (byte-identical output), all overridable by simply setting the variable (DIRECT=0 wins).

Measured on a 9950X3D + Samsung 9100 PRO (Gen5) + RTX 5090, Windows 11 native, 96-token greedy REPLAY, one lever per step, medians of 3–4 runs:

step	tok/s	note
stock	0.45	disk-bound, buffered reads ran ~3 GB/s
+OMP block	0.48	expert-matmul −15%
+DIRECT=1	1.02	+113% (iobench: 10.68 GB/s O_DIRECT vs 9.03 buffered warm; consistent with #162's 1.47×)
+PIPE=1	1.10	workers 4/8/16 flat → default 8 untouched
+PILOT_REAL=1	1.22	hit rate 58→78%

Full ladder (including the negatives: PIN_GB=40 regression, CUDA-tier-alone ≈0% replicating #101, MTP DRAFT=3 losing 0.65 vs 1.05 disk-bound) is written up in the fork's tuning doc: https://github.com/KingIcyCreamProjects/colibri/blob/polish/ui-docs/docs/tuning-9950x3d-5090.md

tests/test_env_defaults.py covers the defaults, explicit-override-wins, kill-switch scope, and the non-win32 no-op. If you'd rather ship these as opt-in (e.g. behind COLI_WIN_TUNE=1) than opt-out, happy to flip the gate — the measurement stands either way.

Note: includes PR 1's commit (physical-core count) as its base — env_for uses physical_cpu_count().

[JustVugg]

Merged to main. Verified: clean merge + all 14 tests pass (test_env_defaults + test_resource_plan). The physical_cpu_count win32 block is identical to #275 (merged just before this) so git folded it cleanly — no duplication. Good call making the I/O defaults setdefault (explicit override always wins) and scoping COLI_NO_OMP_TUNE to the OMP block only; DIRECT/PIPE/PILOT_REAL are all lossless so opt-out defaults are safe. Skipping OMP_PROC_BIND/OMP_PLACES on Windows (MinGW libgomp has no affinity) is the right fix too. Thank you.