Qwen: add gated 0.2.11 acceptance experiments

[atimics]

Summary

This narrows the Qwen work for 0.2.11 to the loader, tokenizer,
configuration, text runtime, diagnostics, a parity-gated optional C recurrence,
and a defensible acceptance runner. Checkpoint-changing work remains explicitly
experimental. PR #30 remains the separate long-term libleCore architecture
track.

The committee proposal and complete v1-v7 provenance are in stacked results PR
atimics/holostuff#3.

V7 result: accepted — GO

Frozen implementation source:
cb3b1d2ac71c183bf9307ca7145a2a619ff30c30.

V7 passed the exact Linux preflight, all twelve ilxyr admission gates, the
experimental installation, all 4,096 paired evaluation positions, checkpoint
reloads, official text generation, the official Transformers vision-input
smoke test, and the mandatory native C qualification. Ilxyr recorded and
independently verified the accepted evidence.

• original and installed perplexity: 7.906224720549084;

• perplexity delta: 0%;

• paired 95% block interval: [0, 0] nats;

• reference-logit relative error: 6.787656774900957e-7;

• reload, generation, vision, tokenizer, statistical, and spectral-disabled
  gates: all passed;

• peak RSS: 19.05 GiB; and

• ilxyr ledger: 16 events and 15 objects, valid: true.

• One-page PDF result

• Committed accepted result

• Permanent evidence

Native C qualification

Both isolated evaluators compiled and used the native GDN recurrence. The
original evaluator recorded 180 native calls over 74,124 tokens; the installed
evaluator recorded 200 calls over 82,360 tokens. Both completed fresh and
resumed-state parity, all float32/float64 checks passed, and neither recorded a
fallback or direct NumPy recurrence call.

The root cause seen in v6 is therefore closed: ilxyr's cleared environment now
receives a minimal deterministic compiler environment, compiler scratch stays
inside the private cache, bounded linker diagnostics are preserved, and CI
fails if an installed compiler silently falls back. The native path remains
opt-in and self-refuses on any parity or provenance failure.

Current implementation

Searchable memory and the boot manifest are SHA-256-bound sidecars; the tied
input embedding/output head remains byte-identical. Weight-resident metadata,
spectral filtering, and checkpoint-changing layer prepending remain
research-only or experimental.

The runner freezes one common paired chunk schedule, permits isolated parallel
evaluators only after memory admission, records per-chunk progress and stage
timings, uses 10,000-draw multiplicity-corrected interim looks for rejection
only, and still requires all 4,096 positions for any GO. Native recurrence is
evaluation-only and parity-gated; compiler, ABI, cache, fallback, call, and
token provenance are recorded.

Performance, cost, and cleanup

Installation took 8,056.60 seconds (about 2h 14m); paired evaluation took
525.35 seconds (about 8m 45s). The installer remains the clear optimization
target. V7 used a 64 GiB r7i.2xlarge, peaked at 19.05 GiB RSS, and cost an
estimated $1.32. The $6.86 launch-tracker figure is the v4-v7 phase
subtotal, not all-history spend: published estimates for v2-v7 total about
$14.23, plus the unestimated v1 attempt and a few cents of storage. The
proposal explicitly says its $10 ceiling buys one formal attempt.

The v7 instance is terminated; its root volume and temporary security group
are deleted. The 95-transaction Arweave publication is confirmed at 0 winc and
verified through the public cache and canonical gateway. No model weights were
published.

Release recommendation

For 0.2.11, retain Qwen compatibility/runtime work, diagnostics, and the
parity-gated native recurrence behind its explicit opt-in. Keep spectral
filtering and weight-resident metadata research-only. Keep layer prepending
experimental until maintainers and the committee decide how the accepted v7
evidence should affect the release gate.

[atimics]

Ready for committee review. The current upstream Actions run is waiting for maintainer approval: https://github.com/AnOversizedMooseWithSocks/leCore/actions/runs/31710930408

Independent validation on the fork is complete:

• strict real-model semantic gate passed unchanged: top-5 8/12, median 1.0, top-1 7/12: https://github.com/atimics/holostuff/actions/runs/31710760534
• normal PR suite passed on the restacked result branch: https://github.com/atimics/holostuff/actions/runs/31711203840
• generated docs passed: https://github.com/atimics/holostuff/actions/runs/31711200705

The separately dispatched exhaustive suite exposes existing 0.2.11 audit debt; its failures reproduce unchanged against the base branch. Please approve the pending upstream run and review the proposed 0.2.11 scope and terminal v2 evidence linked in the PR description.