perf(cpu): isolate Raspberry Pi 5 polling contention

[richiejp]

Row

KERNEL-GEMM-CPU-ELEM-A76 — tracked by #284. The benchmark-control prerequisite is tracked by #293.

Scope

This draft contains the reviewed concurrency-one blocking-c1 benchmark control and the Raspberry Pi 5 W0/C1 records. It preserves poll as the default and rejects blocking-c1 above concurrency one.

It also carries the latest-main integration repairs needed to preserve #206 prompt pretokenization, exact prompt IDs, and atomic ordered wave admission after rebasing.

This PR does not contain the general C2 event wait, the F16 neon-zip / neon-p8 candidates, a default kernel change, or new assembly.

What changed

• Added the explicit poll / blocking-c1 output-wait selector, CLI validation, JSON/text reporting, and exact-token tests.
• Bound the real unconditional get_output_nowait evaluation before the C1 blocking call.
• Preserved current-main pretokenization and wave-admission behavior through the rebase.
• Added mutation-sensitive prompt-ID and checker tests required by the exact PR-size gate.
• Updated STATUS, BENCHMARKS, USAGE, NOW, the owning specs/matrices, and indexed Pi evidence. The llama.cpp floor remains unmet.

Evidence

• Current-head native benchmark suite: 15/15 cases, 124/124 assertions.
• Current-head scoped native CTest: 5/5 (test_engine_core_proc, test_output_processor, test_async_llm, test_bench, test_cpu_threadpool).
• Prompt-ID assignment deletion, nowait bypass, blocking dispatch inversion, report lies, special-token removal, and wave-publication mutations turn RED and restore byte-for-byte.
• Exact PR-size semantic-evidence gate passes in an environment where its sanitized PATH can resolve Git.
• Historical same-binary Pi C1 A/B: 12/12 clean, token-identical, unthrottled. At four compute threads, blocking improved prefill 1.973x, per-stream decode 2.149x, and output throughput 2.151x; T3 was neutral. Four-thread polling was unstable, so this is a causal discriminator rather than a new binding floor.
• Exact-current-head AArch64/QEMU and Pi rebenchmark are owed after the next main rebase.

Review state

• Latest-main C1 conflict integration and its repair passed fresh mutation-driven review. The reviewed repair tree is byte-identical to this PR head.
• Full repository preflight still reports inherited stale live rows and environment/history fixture failures; all change-specific record, document, trailer, registration, PR-size, and scoped test gates pass.

Follow-up work not in this PR

• C2 event wait candidate 518316436: product, native gates, and nine mutations passed review; a record-only anchor correction is required before it can pass final review.
• F16 C++ candidate b120a71b5: native/cross-build/disassembly checks passed; review found that the physical-selection fixture needs an independent correctness golden before Pi benchmarking.
• After those repairs: combine reviewed heads, build under AArch64 QEMU, run same-binary Pi baseline / neon-zip / neon-p8 A/B, retain only a token-exact physical winner, and compare the full model against llama.cpp. Assembly remains forbidden until the C++ path clears that floor.

Rebase status

This branch was rebased through upstream 11d45330b; upstream main is now 6dbedf9f3, and GitHub currently reports merge conflicts. Rebase, fresh conflict review, and exact-head QEMU/Pi evidence remain required before marking the PR ready.

FOLLOWING_AGENTS_PROTOCOL

[localai-bot]

This branch conflicts with main because #364 and #374 removed three shared record surfaces. It is a one-time cost for every branch that was in flight, and the resolution is mechanical — nothing this PR measured or decided needs revisiting. I can't push to your fork, so here is the exact recipe:

git fetch origin main && git merge origin/main

Then, per conflicted file:

File	Resolution	Why
scripts/check-public-doc-tables.py	git checkout origin/main -- scripts/check-public-doc-tables.py	The STATUS_RATCHET["chars"] key is gone. It was a byte count of docs/STATUS.md stored in the checker, so every PR that touched STATUS had to re-pin it. Your re-pin only existed to pay for your own STATUS edit; dropping it loses nothing.
.agents/NOW.md	git checkout origin/main -- .agents/NOW.md	The per-row live-claims table is gone. The live position is derived now — run scripts/now.py. Your row's next step belongs in your row's own spec under a ## Now heading, which has one writer and cannot conflict.
.agents/coordination.md	keep both sides	New claims go in .agents/claims/CLAIM-<ID>.md, one file per claim. The legacy table still parses, so an existing row there is still valid.
.agents/roadmap_v1.md	keep both sides' rows	The issue table is now sorted by row ID. Union both sides and re-sort; do not drop anyone's row.
.agents/benchmark-record.md	union both sides	Append-only log — taking either side alone silently drops a measurement.
tests/scripts/test_check_public_doc_tables.py	git checkout origin/main -- <path>	Follows the checker above.

Anything under src/, include/ or tests/*.cpp is not part of this and is a real conflict worth reading.

Sorry for the churn — removing a shared surface makes every in-flight branch pay once. After this merge your PR stops competing with other PRs for these files entirely, which is the point.