native voice memory: expose owner-scoped high-water accounting and shrink activation/session arenas

[sydneyrenee]

Problem

LfmModelMemoryV1 proves the immutable image and weight-copy invariants, but it does not describe the dominant activation and session allocations. A current one-conversation inventory at the production maximum found approximately 2.281 GiB outside the shared model image, including:

• Conformer workspace: 802,212,288 bytes;
• frontend workspace: 12,312,000 bytes;
• capture-resampled PCM: 1,920,000 bytes;
• mel/adapted rows: 2,304,000 bytes;
• prefill workspace: approximately 5.0 MiB at eight lanes;
• two Flashkern pass scratch banks: approximately 34.7 MiB, dominated by duplicate [32][128000] F32 attention score planes;
• playback PCM: eight separately allocated 1,920-F32 slot payloads;
• capture arena: page-rounded physical storage plus a second virtual alias.

All of this can be legal setup-time storage while still hiding an excessive high-water mark. Cache residency must not be inferred from allocation type: these are ordinary DRAM-backed virtual pages that happen to cache while hot.

Accounting first

Add owner-scoped, versioned native accounting for model, conversation, pass/ticket, and session storage. Report at least:

• persistent recurrence bytes;
• retained formula-derived bytes;
• frontend, Conformer, prefill, Depthformer, sampler, and detokenizer workspace bytes;
• per-pass scratch bytes and slot count;
• capture physical bytes versus virtual alias bytes;
• playback payload and metadata bytes;
• tokenizer/control/event metadata bytes;
• current and maximum live bytes;
• allocation count before readiness;
• post-readiness allocation attempts/bytes.

Accounting must be computed from actual plans/allocations, not hard-coded constants, and must define whether shared storage is charged once or apportioned.

High-water reduction

Use the accounting to drive a liveness plan:

• recalculate Conformer planes from exact producer/last-consumer intervals;
• reuse dead planes only when alias safety is proven at plan construction;
• keep valid mel/Conformer output direct and avoid padded compatibility planes;
• replace prefill and pass-bank vector collections with sealed byte arenas and typed offset views;
• investigate eliminating or tiling the full-context F32 attention score plane without changing trained arithmetic;
• keep ticket-private scratch whenever concurrent passes can still write;
• slab-allocate playback payloads instead of one heap allocation per slot;
• retain the macOS capture double-map as a virtual alias, counting physical bytes only once;
• fuse the final detokenizer → band-limited device-rate output under native inference: remove residual C++ numerical fallbacks and materialized scratch #158 rather than retaining an avoidable full codec plane.

No reduction may depend on assumed L1/L2 residency, macOS affinity, unsafe cross-ticket aliasing, or a scalar fallback.

Verification

• Freeze a baseline manifest for every supported configuration.
• Assert planner high-water equals the sum of actual carved regions.
• Use red-zone/canary and sanitizer tests for every alias boundary.
• Compare full hidden/logit/code/state/PCM values before deleting old layouts.
• Run two simultaneous conversations and both pass slots to prove private scratch remains private.
• Record cold/warm RSS, virtual bytes, physical/accounted bytes, and allocation count.
• Run the native real-checkpoint and audio-only self-conversation gate after each tranche.

Acceptance gates

• Owner-scoped accounting covers every persistent numerical and dock allocation.
• Shared and aliased storage is not double counted.
• No activation or PCM payload uses an unreported heap allocation.
• Conformer, prefill, pass, and playback layouts are sealed before readiness.
• The measured high-water mark decreases without numerical or lifecycle regression.
• No tensor object, per-pass allocation, weight materialization, scalar fallback, or compatibility implementation is introduced.
• AArch64, x86_64/Rosetta, sanitizer, and real-checkpoint gates pass.

Related

• Zero-read/allocation/materialization proof: native voice: enforce zero post-readiness reads, allocations, and weight materialization #157
• Numerical superkernels and materialized scratch: native inference: remove residual C++ numerical fallbacks and materialized scratch #158
• Conversation persistent arena: pending sibling issue
• BlockDomain occupancy: flashkern V2: implement independent four-lane BlockDomains and measured dual-program execution #159
• Release stress matrix: native voice: complete cross-architecture, sanitizer, soak, and fairness release gates #160

[sydneyrenee]

Sibling issue created as #168. #168 owns persistent recurrence/RoPE consolidation; this issue owns measurement plus activation/session high-water reduction.