Spec correction: compression-communication channel definition (Section 1.1) + rewrite of probes 4.22–4.25 + axis-coverage reporting

[FluffyAIcode]

Scope

This PR corrects V331_BLACKBOX_TEST_SPEC.md, not any SUT code. It supersedes an ambiguous reading of 密语系统 / cipher system used in the v3.37 cipher-probe introduction and carried through v3.38-v3.44-Trained.

What was wrong in the pre-v3.45 spec

1. The term 密语系统 was introduced without a precise definition and was interpreted by subsequent audit feedback as "prefix attention alone must carry semantics; any use of content_bias or hard-mask counts as bypassing the cipher." This is not a standard interpretation and it conflated two targets that should have been separate.
2. Three structural probes had unreachable or defective acceptance criteria:
   • 4.23: top-3 of (wte @ slot_1) ∩ rare_keywords >= 1. Qwen 2.5 token ids 0/1/2 (!, ", #) sit near the WTE global mean; any top-K on an unnormalized cosine query is dominated by them regardless of slot content. Metric was measuring WTE geometry, not channel quality. 0/26 pass across v3.38–v3.44-Trained.
   • 4.24: intra_domain_cos_mean - inter_domain_cos_mean >= 0.15 at N=3 memories per domain. JL projection variance into d_ctx=128 is O(1/√N) ≈ 0.58, exceeding the threshold. 0/26 pass across v3.38–v3.44-Trained.
   • 4.25: content_starters_top12_B >= content_starters_top12_A + 1. Saturates at 12/12 in any configuration with a functioning channel; monotone growth impossible by construction. 0/26 pass across v3.38–v3.44-Trained.
3. Anti-cheating clause on 4.22 explicitly banned hard-masking as a solution path. Hard-masking derived from ContentTokenClassifier.pure_function_mask is a legitimate channel mechanism under the corrected definition.
4. Section 4-meta mapped each probe to an attribute from a seven-point "P0..P3 proposals" scheme (声量 / 词汇表 / 抗塌缩 / 调用精细度 / 密语信道容量 / 密语表达形式 / 消歧). This was design-stage motivation, not test semantics.

What this PR adds

Addition	Purpose
Section 1.1	Precise four-axis definition: A compression ratio, B O(1)-in-N injection cost, C semantic fidelity vs naive RAG, D channel stability
1.1.2	Explicit enumeration of legitimate channel mechanisms (prefix, content_bias, suppression_bias, fwd_function_suppression, mean-centered residual, retrieval ranking, context_descriptor)
1.1.3	Narrower ban list: prompt-keyed routing, mocks, memorised templates, audit-only code paths, stub backbones
1.1.4	Historical note on the v3.38-v3.44 probe misspecification
4.23 correction	Mean-centered top-20 intersection + median rank_of_best_rare <= 100 out of vocab 151936. Passable by an actually-functioning tail subchannel.
4.24 correction	LOO NN classification accuracy >= 0.75 at N=8. Clopper-Pearson CI is bounded; JL variance no longer dominates.
4.25 correction	Starter-mass ratio mass_B / mass_A > 1.10. Unbounded above, monotone in capacity, not saturation-bound.
4.22 correction	Remove hard-masking exclusion. Metric and threshold retained.
4.21, 4.20, 4.26 correction	Axis re-mapping only, no metric changes
4-meta rewrite	A/B/C/D axes replace seven-point attribute scheme. Gating table shows pre-v3.45 vs v3.45+ distinction.
4-meta.1	NEW — axis-coverage table required in every v3.45+ audit report
Section 7.7	Channel-axis framing rules for human-authored reports. Bans value-judgment use of "cipher works"; requires axis-specific numeric claims instead.
Section 7.8	Retraction notice: pre-v3.45 feedback documents that claimed a single probe's PASS or FAIL established or refuted channel existence are superseded.

What this PR does NOT do

• Does not modify v331_blackbox_eval.py runner code. The runner will be updated in a separate PR once the spec has landed.
• Does not rewrite pre-v3.45 audit feedback documents. Retroactive retraction is handled by Section 7.8.
• Does not change any SUT (scheme_b_vXXX.py).
• Does not change pass/fail gating of any Section 4.1–4.19 case.

Runner update requirements (for a follow-up PR, not this one)

To implement the corrected 4.23/4.24/4.25 metrics, the runner must:

• compute wte_mean once at runner startup and pass it to the 4.23 probe
• add sklearn-style LOO NN implementation (or a 20-line numpy equivalent) to the 4.24 probe
• compute baseline-shifted starter mass in the 4.25 probe
• emit the Section 4-meta.1 axis-coverage table at the end of report.md

Until that follow-up lands, the v3.45+ audit runs with the old runner will still fail 4.23/4.24/4.25 on the pre-correction metrics. The spec correction is valid independently; it defines the target that the runner update must meet.