== ADDENDUM 2026-09-15 (lia) — READ BEFORE THE BODY BELOW ==
The brief below is a 2026-09-12 snapshot. Two things changed; where they conflict, this wins.

1. RUN 7 IS LOGGED AND HAS AN OPERATOR VERDICT. The section headed "THE OPEN ITEM I FOUND TODAY"
   is superseded — do not go ask Reavo for that verdict, it exists. The 2026-09-10 12:45-14:29 run
   is now written up in M3-PLAN.md as "M3 run 7 — operator verdict recorded", with Reavo's ear:
   word endings seem okay (provisional), loudness okay for this run, BUT reverb is still too wet,
   and Local Voice sounds over-processed and robotic next to other players, while the raw mic feed
   does not. That robotic coloration is an open unattributed symptom — not yet pinned to reverb,
   codec, dynamics, pitch processing, or output monitoring. It is now T4a's job to diagnose it,
   alongside the reverb parameter matching. Run 7 signs off neither T4e nor the megaphone: bell
   bloom openings 0, pitch 1.000, speechlessness 0.00, broadcasts 0.

2. REPO STATE AS OF 2026-09-15, NOT AS WRITTEN BELOW. HEAD is still 17fb547 on main, in sync with
   origin/main — but the tree is NOT clean. Two files are modified and UNCOMMITTED: M3-PLAN.md
   (the run 7 write-up above, +44 lines) and docs/RELEASE-RUNBOOK.md (fresh-install check corrected
   to match T4d — it no longer tells you to look for a Settings > Audio row that T4d deleted, and
   now has you confirm the Helper's offset caption instead). This work has been sitting uncommitted
   since roughly 2026-09-12. First thing in your seat: review those two diffs and commit them
   before anything else touches the repo, so the verdict is not lost to a stray checkout.

Net effect on the ordered next steps below: step 1 (a run that actually exercises bells and the
megaphone) is unchanged and still the gate. Step 2 (T4a) grows a third job — diagnose the robotic
coloration — and the brief's note that T4a "settles the loudness question" is now secondary, since
the operator called loudness okay on run 7; wetness and coloration are the live fidelity problems.
Keep the same signal's decoded input.wav, local-voice.wav and the Switch listener capture so the
processing delta can be separated from the peer reference, plus a raw-mic recording as the clean
baseline.