Issue #117 Arm-C SWA allocation remediation (InferSwarm #166): lifecycle-owned full->swa mapping in the standalone R6 stage runtime

[ezutfen]

Closes InferSwarm FlashML-org#166 (implementation remediation only; NO Arm-C requalification).

Starting heads (verified exact at session start)

• FreeToken base: 55e8baaebabe67aeb967d4bd407ef26696933104 (accepted research head)
• InferSwarm base: df0365ea606a54222415e34f6cfb127e9f938f0c (main, Merge PR fix(kernel): GGUF JIT extension fails to build with gcc under C++17 - compile as C++20 FlashML-org/FreeToken#159)

The defect (FlashML-org#157-localized)

The standalone R6 stage path builds the same HybridSWAKVCache pool as the scheduler but never reaches the scheduler's alloc_swa lifecycle (CacheManager.allocate_paged): full_to_swa_index_mapping stayed at the all-zero sentinel, every SWA-layer KV store raced on swa slot 0, and every prefix read consumed slot-0 bytes (chunk-2 nondeterminism; earliest varying boundary L0_kv_slice_post_write).

The remediation

GemmaDenseStage owns the session SWA allocation lifecycle at the same lifecycle points the scheduler uses:

• prefill()/decode() (all four role branches) call _ensure_swa_session_mapping BEFORE _prepare (prepare_metadata translates the full page-table range through the mapping — the first prefix-consuming SWA operation of each chunk);
• incremental ownership frontier [0, allocated) matching CacheManager.allocate_paged's per-chunk granularity; each position's slot allocated exactly once; mapping persists across chunks and decode;
• reset_session_state rebuilds the allocator AFTER the zero loop (the zero loop wipes _swa_free to all-zeros, which would hand out sentinel slot 0 on the next allocation — the Feat/sm75 cuda128 FlashML-org/FreeToken#157-recorded trap), releasing all ownership wholesale: reuse cannot inherit stale mapping;
• exhaustion and position gaps fail CLOSED (RuntimeError) — execution never continues on the sentinel;
• non-SWA pools untouched no-ops; pool sized max_seq_len+1 (sentinel convention) so a full-capacity session fits.

No parallel allocator (existing pool primitives reused verbatim); no case/regime/model special cases; 64-row boundary contract, geometry, planner, tokenizer, thresholds unchanged.

Validation (CPU-only; no GPU/model execution)

• NEW tests/research/test_issue166_swa_remediation.py: 35 lifecycle tests incl. negative controls (skipped allocation, delayed allocation, slot-0 aliasing, stale-mapping-survives-reset, ownership leak, ignored exhaustion, case-gating)
• Focused regression sweep 189/189 green (issue117_arm_c_remediation, r6_gate_contracts, r6_final_row_fp32_promotion, r6_inplace_softcap, r6_single_gpu_control, r6_single_gpu_control_gate)
• tests/research/ 563 passed; the single failure (test_r6_localization::test_stage_chain_capture_ops_are_explicit) is proven pre-existing at base 55e8baa (stash-and-rerun); scheduler/kvcache SWA suite failures likewise byte-identical pre-existing on this CPU-only host
• Causal applicability: the lifecycle's alloc_swa(arange(frontier, need_end)) produces the byte-identical mapping the accepted Feat/sm75 cuda128 FlashML-org/FreeToken#157 intervention produced (test_standalone_reaches_the_scheduler_allocation_semantic)

Boundary

Terminal: ISSUE117_ARM_C_SWA_REMEDIATION_READY (InferSwarm record PR). This producer is NOT promoted to qualification authority; fresh Arm-C requalification remains separately blocked pending maintainer acceptance/merge. InferSwarm-side record: Zutfen-LLC/inferswarm branch issue-166-swa-remediation @ e13cf58.