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, and diagnostic surface, while keeping checkpoint-changing work explicitly experimental.

The committee-facing proposal and terminal v1 execution record are in stacked draft PR atimics/holostuff#3. Review implementation here and experiment policy/evidence there.

• disables spectral filtering by default and labels it research-only
• requires --experimental for the layer-prepending installer
• fixes paired statistical inference and the preregistered non-inferiority gate
• removes avoidable float64 amplification and uses file-backed lazy loading
• fixes the caller-relative shell installer path
• separates installation and held-out evaluation corpora and rejects identical content
• requires clean source, tokenizer parity, and reference-logit parity before installation
• records a complete top-level model-file manifest
• preserves the requested virtual-environment interpreter instead of resolving its symlink to the system Python
• replaces the machine-specific fixture with a portable, offline miniature Qwen fixture and CI contract

Acceptance contract

The real-model contract requires:

• exact tokenizer IDs and reference-logit parity before installation
• at least 4,096 paired evaluation positions
• a 95% paired moving-block bootstrap and preregistered +1% perplexity-regression ceiling
• peak process and accelerator memory reporting
• reload and finite-logit execution of the emitted checkpoint
• text generation through official Transformers
• an official Transformers image-input generation smoke test
• spectral filtering disabled throughout

First real execution

The owner-authorized v1 AWS attempt is complete with ilxyr outcome execution_failure. It passed all frozen-input checks and all twelve ilxyr admission gates, then exited after 51 ms before model evaluation because the generated command invoked /usr/bin/python3.12 rather than the prepared virtual environment. No scientific acceptance/rejection claim can be made from v1.

That concrete runtime bug is fixed by 0bda4f1, with a regression test. The full result, verified .ilxyr ledger, native/RO-Crate/in-toto exports, raw logs, and receipt are committed in results PR #3 and permanently indexed at Arweave root jguW…mHBk.

This PR does not claim that Qwen3.5-0.8B has completed the acceptance contract. A corrected v2 run needs a new experiment identity and explicit authorization. The installer remains experimental.

Validation

• 10 passed in tests/test_qwen_acceptance.py
• portable Qwen fixture loads, tokenizes, and runs from an arbitrary working directory
• experimental miniature-checkpoint installation emits and reloads its output
• generated project exercises ilxyr contribution, compile, forecasts, funding, admission, execution record, and workspace verification
• v1 ilxyr ledger verifies: 15 objects, 16 events, valid: true
• v1 permanent publication: 63 confirmed Arweave transactions, zero winc charged, public re-download hashes matched
• focused Qwen and wiring regression suite before launch: 22 passed
• UNICRON safetensors/GGUF round-trip self-test: passed
• generated-doc and build-ergonomics regression suite: 12 passed
• hosted Qwen contract and aggregate gates previously passed in run 31627998803

Release gate

For 0.2.11, retain the compatibility/runtime work and diagnostics. Keep spectral filtering research-only and the layer-prepending installer behind its experimental acknowledgement. Do not promote the installer until a new complete real-model acceptance run reaches and passes every gate.

PR #30 remains the separate long-term liblecore architecture track.