v3.45 cond-buffer: fix 4.23 aliasing, audit 21/26 (+1, no regressions)

[FluffyAIcode]

What this PR does

Fixes the pre-existing diagnostic-buffer aliasing in EmbBridge that caused case 4.23 keyword_specific_tail_slot_probe to read uncond-path tail slots instead of cond-path tail slots. No Cfg change. No algorithm change. No SPEC change. One SUT-side data structure extension; one runner-side read-site change; one diagnostic script.

Root cause

MemLLM.prepare_decode_context calls EmbBridge.inject twice in the CFG decoding path:

1. Cond — with rare_keyword_wte_residual=residual, produces prefix_cond.
2. Uncond (_build_contrastive_uncond_prefix) — with rare_keyword_wte_residual=None, produces prefix_uncond.

Both writes go to bridge._last_tail_slots, _last_residual, etc. 4.23 reads bridge._last_tail_slots after prepare_decode_context returns, so it always saw the uncond prefix's tail slots — which by construction carry no rare-keyword signal.

Observed on v3.48 and v3.45: top-5 = [' ', ',', '.', ' (', '1'], median_rank ≥ 1089 across both versions. The number was noise from the uncond inject, not a measurement of the cond tail.

Fix

SUT (scheme_b_v344.py)

• EmbBridge.__init__ adds six _last_cond_* mirrors.
• EmbBridge.inject signature gains is_cond_path: bool = True.
• EmbBridge.inject epilogue: when is_cond_path=True, copy each _last_* into the matching _last_cond_*. When False, only the shared _last_* are written.
• MemLLM._build_contrastive_uncond_prefix passes is_cond_path=False on its inject call. All other call sites default to True.

Runner (v331_blackbox_eval.py)

• keyword_specific_tail_slot_probe prefers bridge._last_cond_tail_slots via _get_tail_slots_cond_preferred(model), falling back to bridge._last_tail_slots for older SUTs.
• Return payload records tail_slots_source.
• metric_version → v3.50.

No:

• Cfg change
• algorithm change (residual composition, attention pool, inter-domain margin, top-1 bias, circuit breaker — all untouched)
• SPEC change (primary metric formula identical; only which tensor is read)
• training change

Pre-audit diagnostic

diag_4_23_cond_buffer.py compares _last_tail_slots vs _last_cond_tail_slots after prepare_decode_context:

query: She performed Beethoven sonatas with delicate phrasing on her grand piano.
  dom_mem rare_keyword_ids = [2524, 14317, 14762, 26278, 37191, 38744, 48601]
  pieces=[' control', ' finger', ' technique', ' piano', ' refined']

  _last_tail_slots (shared):     slot_1 L2=0.0000   top5=[' ', ',', '.', ' (', '1']
  _last_cond_tail_slots (new):   slot_1 L2=1.0251   top5=[' control', ' Control', '控制', 'control', 'Control']
  rank of 'control' (id=2524):   was 1402   now 1

Matching result for the second paraphrase.

Audit result

21/26 pass in 1547 s under the v3.49 runner on fresh-init, CPU, AMS_DETERMINISTIC=1.

case	v3.48	v3.44-rewrite	v3.45 (revertB+refreshD)	v3.45-condbuffer
4.23 keyword_specific_tail_slot_probe	FAIL	FAIL	FAIL	PASS

4.23 primary metric numbers:

tail_slots_source = bridge._last_cond_tail_slots
mean_intersection_size_top20_paraphrase = 1.0     (threshold >= 1.0)
median_rank_of_best_rare_paraphrase     = 1.0     (threshold <= 100)
hit_ratio_at_least_one_top20_paraphrase = 1.0     (threshold >= 0.5)
n_paraphrase_queries_evaluated = 2

Every other case holds its state. No regressions, no NI flips.

Axis coverage

axis	metric	status
A compression	1712 floats / (10 × 1536) = 8.97 (threshold 10.0)	FAIL
B injection cost	164224 per-step, O(1) in N	PASS
C fidelity	8/11 (was 7/11; 4.23 added)	FAIL (threshold 9)
D stability	2/3 (4.21 FAIL)	FAIL

Causal chain validated end-to-end

The v3.45-revertB-refreshD branch identified the aliasing bug and left the question open: with the bug fixed, does the reverted additive path (slot_1 = tail_head(fiber) + α·residual) actually carry the rare-keyword signal? Answer from this audit: yes. On fresh init, tail_head.slot_heads[1] is zero-init (per tail_head_zero_init_tied=True), so slot_1 = α·residual by construction, which at α=1.5 and residual = strict-keyword WTE centroid has the rare keywords as its top-k argmax directly. The v3.44-rewrite's combine_with_residual (β·LN(head_out) dominating at L2=11.76) replaced this clean geometry with an LN-gamma direction — the plan's magnitude argument is now numerically confirmed.

Remaining FAILs (untouched)

• 4.7 semantic_memory_counterfactual_pairs — repetition garbage
• 4.8 degeneration_quality — same repetition root
• 4.11 retrieval_topk_semantic_shift — prefix → meta-starter mismatch
• 4.19 stepwise_label_mass_alignment_audit — cascade of 4.11
• 4.21 decode_repetition_feedback_probe — same repetition root as 4.7/4.8

These are the cases plan #2 (narrow [E]) and #5 (rare_keyword floor) target, and they are orthogonal to this PR.

Next decision point

With 4.23 now measurable and passing under #1+#3+cond-buffer, two remaining failure clusters:

Cluster X (decoder repetition, 3 cases): 4.7 / 4.8 / 4.21. Root cause identified in v3.44-rewrite post-mortem: top1_content_bias_weight=0.7 × scale=6.0 × unit=3.75 × relevance=0.79 ≈ +22 logit outpaces content_repeat_penalty=2.5 × k until k≥10, while cyclic_content_max_count=5 hard-masks at step 5. Plan #2 narrows scales to let penalty win at k=3.

Cluster Y (prefix lexical class, 2 cases): 4.11 / 4.19. Root cause: prefix → Qwen 28-layer forward → final hidden has ~0.14 logit contribution in any rare-keyword direction, below the ~10 logit baseline of modal transition tokens. Plan #5 adds rare_keyword_floor_boost on the fwd-path content_bias to bypass this.

Awaiting direction on which cluster to address first, or whether to stop here.