Add 3x9 factorial analysis of std79-doe heartbeat/physics telemetry

Correlates every HB-ON/HB-OFF heartbeat-tick CSV row (results-20260912-
with-heartbeat-csv/*-doe-raw.log) back to which trial (run_id, id_idx,
id_label, rep) was active when it printed, despite the async tick
printer splicing rows mid-token -- including mid a DOE-RUN marker itself
-- into the trial loop's own console output on the shared serial line.

Pipeline (analysis-20260912/, see its own README.md):
- correlate_doe.py: two-pass reconstruction per architecture (remove
  atomic CSV-row spans to rebuild the clean trial-output stream, map
  each removed row's offset back to the nearest preceding run_id
  marker); identity/rep looked up from a known-clean prior run's
  run_id mapping rather than re-parsed, since one aarch64 marker
  (trial 12, rajames rep 1) lost its id_idx digit to a zero-separator
  collision with an adjacent CSV field and is unrecoverable from that
  log alone -- its rows fold into trial 11 instead, documented as a
  known limitation.
- combine.py: merges all three architectures into combined.csv (768
  rows), decoding Q48.16 fields to floats and jitter_bits' IEEE754 bit
  pattern to real jitter_ns.
- analysis.R: per-cell (architecture x identity) means/SD and two-way
  ANOVA for each of 12 telemetry metrics, one boxplot SVG per metric,
  written up as ANALYSIS.md.

Key findings: identity significantly affects word-heat/window-sizing
metrics (expected -- different identities execute different word
sets), architecture significantly affects timing metrics (APIC
ticks/tick, timing variance, fleet heat -- expected, different QEMU
targets), zero architecture x identity interaction on any metric.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EXieurDfDSsDFdnSyusuWo

experiments/std79-doe/README.md

@@ -68,3 +68,15 @@ nothing is corrupted in memory — only the printed character stream interleaves
 tooling that wants an exact reconstruction of the trial-output stream from these three raw logs
 needs to account for that (strip `[HADES][DOE ] ...` fragments and rejoin) rather than assume
 one physical line is one logical print.
+
+`analysis-20260912/` builds exactly that reconstruction and turns it into a real 3×9 factorial
+analysis: `correlate_doe.py` labels every heartbeat-tick row with the trial that was active when
+it printed, `combine.py` merges all three architectures into `combined.csv` (768 rows, Q48.16
+fields decoded to floats), and `analysis.R` computes per-cell (architecture × identity) means/SD
+and a two-way ANOVA for each of 12 telemetry metrics, rendering one boxplot SVG per metric. See
+`analysis-20260912/README.md` for the pipeline and `analysis-20260912/ANALYSIS.md` for the
+report itself — key findings: identity significantly affects word-heat/window-sizing metrics
+(expected — different identities execute different word sets), architecture significantly
+affects timing metrics (APIC ticks/tick, timing variance, fleet heat — expected, different QEMU
+targets), and **zero architecture × identity interaction on any metric** — the two factors'
+effects are clean and separable, not confounded.