v3.41 black-box audit: G-1..G-7 structural fixes for 10 v3.40 failures

[FluffyAIcode]

Scope

• SUT: scheme_b_v341.py via AgentMemorySystem.py redirect.
• Runner: v331_blackbox_eval.py, unmodified.
• Spec: V331_BLACKBOX_TEST_SPEC.md, unmodified.

v3.41 architectural changes

Tag	Target cases	Mechanism
G-1	4.17	MemoryContextEncoder fp32 + no mean-center; save uses .detach().contiguous().cpu()
G-2	4.10, 4.22	_get_prefix sets _mem_guidance_active and attaches content_bias/suppression_bias in both return_extra branches
G-3	4.12, 4.23	rare_keyword_wte_residual at native WTE scale; blended post-aligner at α=0.5
G-4	4.24	MemoryContextEncoder.encode_from_tokens uses strict-starter WTE centroid
G-5	4.25	ContentSemanticTailHead tied_extra=True (2 distinct heads regardless of L_mem)
G-6	4.15, 4.22	fwd-path function suppression decoupled from fwd_path_bias_dampen
G-7	encoder training	Trainer.context_separation_loss (warmup-gated)

Audit result

• 26 cases, elapsed 1437.7 s on CPU.
• Pass: 17 / 26, Fail: 9 / 26.
• v3.40 baseline: 16 / 26 pass, 1309.4 s.

Delta vs v3.40:

Transition	Count	Cases
FAIL → PASS	3	4.6, 4.12, 4.22
PASS → FAIL	2	4.8, 4.21
Persistent FAIL	7	4.7, 4.10, 4.15, 4.17, 4.23, 4.24, 4.25

The two new regressions (4.8, 4.21) share a single mechanism: [G-2] attaching content_bias to prefix_cond and [G-6] decoupling function-suppression produces an effective double-add of content bias (once in fwd, once in shape_step_logits), driving early token repetition. Falsifiable check documented in audit_feedback.md §4.3.

Artifacts

• reports/v341_blackbox/report.json
• reports/v341_blackbox/report.md
• reports/v341_blackbox/runner.log
• reports/v341_blackbox/audit_feedback.md (Section 7 compliant)