[novkien]

Addresses #9

Overview

When pipeline copy-buffer reservation fails, retry with two copies before the existing one-copy generic fallback. Successful four-copy reservations stay unchanged; this does not set a global two-copy default. The Flash acceptance checks reject the one-copy result; generic one-copy fallback remains available.

Validation

• Focused test-alloc passed on CPU and on the matched native/RPC build.
• Matched source 109ab654 plus this patch passed a proxy :8082 capacity screen with RPC-first placement, ub=1024, two slots, pipeline enabled and all 49 target plus 50 draft layers on GPU. At shared contexts 233472, 249856 and 266240, the initial four-copy reservation failed and the two-copy retry succeeded. One cold 8192+512 request completed at each context.
• PP/TG were 492.56/69.69, 503.59/71.43, and 485.60/70.01 tokens/s at those contexts, respectively. These are single cold requests per context, not a performance qualification.

Remaining qualification

This PR does not claim maximum-context or near-full-pool acceptance. Long-context behavior, cancellation, simultaneous two-slot qualification, combined-route qualification and final deployment remain outstanding in issue #9. Root will continue this qualification after merge; this PR is not final task acceptance.

Requirements

• I have read and agree with the contributing guidelines.
• AI usage disclosure: YES - Codex authored and independently reviewed the implementation and tests under the owner-approved plan; this submission is AI-assisted.

[novkien]

Pre-merge handoff to Codex /root: this issue-linked draft is at reviewed head 68136f094fcec0d898eb037552dc1f7af621fbca on base 109ab6544090ba9922496c8ca4b34fe6751120ca. The issue #9 publication report records the scope and evidence.

Focused test-alloc passed on CPU and the matched native/RPC builds. The proxy :8082 capacity screen retained all target/draft layers on GPU with two slots and pipeline enabled. It passed one cold 8192+512 request at each of c=233472, 249856 and 266240 after four-copy reservation failed and two-copy retry succeeded. This is capacity evidence only.

Maximum-context and near-full-pool acceptance, long-context behavior, cancellation, simultaneous two-slot qualification, combined-route qualification and final deployment remain outstanding. PR remains draft; merge is on hold for root's continuing qualification.

[novkien]

Post-merge handoff: PR #20 is squash-merged at c09fa602cc64e30eef5db0cb3784667a91172452, from reviewed head 68136f094fcec0d898eb037552dc1f7af621fbca. Issue #9 remains OPEN; the post-merge issue report records the remaining acceptance work.

Root will continue exact merged-source deployment and proxy verification, maximum-context/near-full-pool, long-context, cancellation, simultaneous two-slot and combined-route qualification after merge. The combined trace reported by root (70,000 new tokens plus 512 generated, GPU-resident layers, two-copy pipeline) is state/route evidence only. No final acceptance or deployment is claimed here.