feat: harden measurement + scoring pipeline (audit A1/A3/A4)

[BlueBirdBack]

Summary

Hardens the measurement and scoring pipeline per the source-audit findings (A1, A3, A4):

• Memory test measures DRAM, not a shared cache buffer: --memory-scope=local gives each sysbench thread its own working set, auto-sized as RAM/threads/4 clamped to [32M, 256M] (overridable via MEM_BLOCK). The previous --memory-block-size=1M + default global scope benchmarked a 1 MiB hot buffer.
• Parse failures can no longer masquerade as zeros: malformed/jobless fio JSON and missing p95 metrics report FAILED and degrade the run to PARTIAL, instead of scoring a fake 0.00.
• --strict flag / ABS_STRICT=1: exit non-zero when the verdict is INCOMPLETE, so bash abs.sh && ... is trustworthy.
• Command timeouts: sysbench/fio calls wrapped in a portable watchdog (ABS_CMD_TIMEOUT, default 900 s) so a hung benchmark cannot stall the script forever.
• Reproducible JSON: new provenance block records script SHA-256, sysbench/fio/python3 versions, and the score_model id; mode now includes cpu_prime, mem_block, strict.
• Regression tests: coverage for per-thread scope selection, malformed fio JSON → FAILED, and --strict exit codes. 22/22 pass, shellcheck clean.
• Scoring reads structured metrics.tsv, not display text: every measured value is captured as a numeric field at add() time; all three scorers (local score, network score, verdict) consult it first and only fall back to text parsing — the whole A2-class of "digit inside a label gets scored" bugs is eliminated, not just its instance.
• All benchmark calls watchdoged: fio (including --end_fsync=1 prepare), iperf3 (portable abs_timeout, no GNU-timeout dependency), and the dd fallback are now time-bounded.
• Scoring model doc (docs/scoring.md): abs-local-v1 formulas, the provenance of every reference constant, verdict thresholds, comparability contract, and honest calibration limits — linked from both READMEs.

Verified end-to-end on macOS (Apple Silicon): full run completes, memory now reports ~79 GiB/s per-thread DRAM bandwidth (was ~183 GiB/s cache-hot), FULL score 2532.

Notes

• This branch contains the commit from fix: require numeric EPS/throughput fields before scoring sysbench results #2 (EPS field validation) — merge fix: require numeric EPS/throughput fields before scoring sysbench results #2 first, or review both together here.
• CI macOS job could not be pushed (PAT lacks workflow scope). Apply this patch to enable a ubuntu+macOS test matrix:

--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -9,13 +9,23 @@
 jobs:
   regression:
-    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+    runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v5
-      - name: Check shell
+      - name: Check shell (Linux)
+        if: runner.os == 'Linux'
         run: |
           sudo apt-get update
           sudo apt-get install -y shellcheck
           shellcheck abs.sh
+      - name: Check shell (macOS)
+        if: runner.os == 'macOS'
+        run: |
+          brew install shellcheck
+          shellcheck abs.sh
       - name: Run regression suite
         run: python3 -m unittest -v tests/test_abs.py

Devin session: https://app.devin.ai/sessions/c74df4e507504d63aa16cf898640ae58
Requested by: BlueBirdBack