Pre-compact cleanup: CHANGELOG Sprint 7 + archive entry + memory

Sprint 7 covers the morning: mi-UGens s.sync, the preload pointed at what gets
played, three stacked bugs in its freshness check, 198 AppleDouble deletions,
s.latency 0.3->0.2, and the wireplumber start-limit outage.

Archive entry carries nine learnings, the load-bearing ones being that pgrep
costs 53-59ms a call here while the real trick is the cheap prefilter (the regex
on every comm is WORSE than the pgrep it replaced); that case-in-a-variable does
no alternation; that start-limit-hit is this box's signature trap and retrying is
the one action that cannot work; and that a valid measurement can be silent,
because mute sits downstream of all DSP while a null sink lies.

Memory: verify-the-leg-that-breaks gains the three 2026-09-07 instances plus two
rules (ask why two numbers disagree; a tool owns the commands it prints).
rig-state records the reset-failed lesson and the BT-by-name detail.

CHANGELOG.md

@@ -4,6 +4,69 @@ Sprint entries, newest first. Player-facing: what changed about *playing*, not
 about the code. Task IDs reference the L'Armada board; `n/a` where the work was
 unplanned (which, on a gig night, is most of it).
 
+## Sprint 7 — 2026-09-07 (morning) · the rig stops lying about itself
+
+Five separate things were reporting green while doing nothing. None of them were
+found by reading code; every one came from making the rig say out loud what it
+was actually doing, then listening to the part that did not match.
+
+### Fixed
+- **Every orbit's MiVerb / MiClouds / MiRipples send was a dead node** (n/a).
+  84 `SynthDef global_mi_verb2 not found` lines per boot — 3 defs × 14 orbits ×
+  2 — so `# verbwet`, `# cloudswet`, `# ripplesreson` did nothing, on any orbit,
+  all session, silently. mi-UGens was installed perfectly and the SynthDefs were
+  written correctly; `.add` is asynchronous and the next statement asked the
+  server for those names. One `s.sync`. 84 → 0.
+- **The preload warms what gets PLAYED** (n/a) — 60 banks → 124, 16 tracks → 39,
+  computed from the canonical gig records rather than a hand-kept file. 51 banks
+  were one first-play from a disk read.
+- **The preload freshness check is trustable** (n/a) — three stacked bugs: a
+  headline that sent you to regenerate when regenerating could not help; a
+  comparison of bank NAMES that never looked at counts (so a new pack dropped
+  into an existing bank was invisible under a green check); and a case-blind
+  regex that truncated `rampleA0` to `rample` and collapsed the whole `vocal*`
+  family, which is why the checker said 111 while SuperDirt loaded 124.
+- **198 AppleDouble resource forks deleted** (n/a), on PLN's say-so, every one
+  confirmed by `file` magic rather than by name. They failed to load AND
+  occupied sample indices in four `rhadamanthe_*` banks.
+- **`s.latency` 0.3 → 0.2** (n/a). The comment said "increase this if you get
+  late messages" and the rig had 136 — but they came in two bursts with values
+  decaying 8.1 s → 0.65 s, which is a backlog draining, and both sat against
+  restarts. The reason to pad it was a misdiagnosis of a bug now fixed.
+- **All audio came back** (n/a). Reported as dead media keys and a silent
+  Bluetooth headset; actually `wireplumber` in `failed (start-limit-hit)` after
+  five hand restarts in two minutes, leaving `Dummy Output` as the only sink.
+  The headset itself was merely muted at 20% on an already-correct `a2dp-sink`
+  profile.
+
+### Added
+- **`tools/gen_setlist.py`** (n/a) — computes what was played in the last N
+  months from `lives/*/tracks.json` plus the ear-derived `judge_specs`, so the
+  preload's input cannot go stale. Ear lists carry no gig date, so it includes
+  them and SAYS SO rather than dating them by guesswork.
+- **`rig-doctor` gained three checks that ask about function** (n/a):
+  SuperDirt boot errors (scoped to the current boot, names the missing
+  SynthDefs), the mi-UGens plugin/class inventory, and the audio session
+  (wireplumber up AND a real `alsa_output`/`bluez_output` in the graph, never
+  `auto_null`). The last one caught a live regression the first time it ran.
+  47 → 48 checks, 0 fail.
+
+### Doctrine
+- **`start-limit-hit` is this box's signature trap, and the instinct is the one
+  thing that cannot work.** Three units in one session — `parvagues-sc` twice,
+  `wireplumber` once — left permanently unstartable, each after someone
+  restarted them again. "X will not start" means `reset-failed` FIRST.
+- **A test must not grep for a word its own good news contains**, and an error
+  is not a negative: the mute guard read a wireplumber `Translate ID error` as
+  "unmuted" and aborted a six-minute sweep naming a cause that never happened.
+- **Read the exit code of the command, not the pipe.** `check-preload.sh` was
+  briefly accused of exiting 0 while reporting STALE; the 0 was `tail`'s. Same
+  family as `grep -c` printing 0 and exiting 1.
+- **Advice is part of the tool.** `latency-lens --print-config` handed over
+  `systemctl --user restart pipewire pipewire-pulse wireplumber` with no
+  warning, which repeated is exactly the trap above. A tool that prints a
+  dangerous command owns the consequence.
+
 ## Sprint 6 — 2026-09-07 · the crackle that was waiting at the venue
 
 Went looking for latency to shave. Found that the thing making the rig miss its
@@ -28,7 +91,7 @@ generated on this laptop.
 
 ### Found, not yet fixed
 - **The plan is now fresh and pointed at the wrong set** (n/a). Re-running the
-  identical load after the fix still logged 29 lazy reads — and every one was
+  identical load after the fix still logged 41 lazy reads — and every one was
   the `rose` bank, because `rose_rouge` is not in `setlist_opal2026.txt`. Zero
   of the 60 preloaded banks lazy-loaded, so the mechanism is provably working:
   it is warming the 16-track OPAL set while the track this box actually plays

armada/tasks/completed-archive.md

@@ -7,6 +7,69 @@ snapshot in `board-archive.md`. Newest first.
 
 ---
 
+## n/a — the rig stops lying about itself
+
+**Description.** "Fix parvagues-protect burning 10.7% of a core, then do a pass
+on perf and tooling efficiency/footprint." Which turned into: find out why the
+box misses its deadlines, and fix everything that was reporting green while
+doing nothing.
+
+**Done.**
+- `parvagues-protect` 10.96% → 1.42% of a core, 6 forks/tick → **0**. All of the
+  219 ms was discovery, none protection: three `pgrep -x` at 59 ms each.
+- `sc-watchdog` no longer restarts SuperDirt on every clean start. The unit goes
+  `active` when sclang execs; scsynth appears ~8 s later; the loop acted at 6 s.
+  Each clean start spent two of systemd's three allowed starts.
+- `preload.scd` generated, then pointed at what actually gets played: 60 → 124
+  banks, 16 → 39 tracks, computed from the gig records by `gen_setlist.py`.
+- Three stacked bugs in the preload freshness check; three numbers now agree.
+- 84 `SynthDef not found` errors per boot → 0, with one `s.sync`.
+- 198 AppleDouble resource forks deleted, confirmed by magic not by name.
+- `s.latency` 0.3 → 0.2; `late` is now a counted metric that rejects a quantum.
+- All audio restored from `wireplumber: failed (start-limit-hit)`.
+- rig-doctor 45 → 48 checks, 0 fail, and three of them now ask about function.
+
+**Learnings.**
+1. **`pgrep` is not free.** 53-59 ms per call on this box, because it reads
+   `cmdline` for every process. One pure-bash pass over `/proc/*/comm` is 21-28
+   ms and forks nothing. But the win came from the *prefilter*: a `[sSaA]*`
+   character class rejects 606 of 641 processes before bash's regex engine,
+   which is the slowest thing in reach — testing the regex on every comm costs
+   149 ms, **worse than the pgrep it replaced.** Measure the replacement too.
+2. **`case "$x" in $pattern)` does not do alternation.** `|` is recognised when
+   `case` is parsed, so `s*|S*|a*|A*` arriving from a variable is one pattern of
+   literal pipes and matches nothing. Every part was individually correct and
+   the scan came back empty. A bracket class survives expansion.
+3. **`start-limit-hit` is the trap, and retrying is the one thing that cannot
+   work.** Three units, one session. `reset-failed` first, always.
+4. **The tool's advice is part of the tool.** Printing
+   `systemctl --user restart pipewire pipewire-pulse wireplumber` without the
+   start-limit warning is handing over a loaded gun.
+5. **A generated artifact stays fresh only via the one path nobody took.**
+   `preload.scd` was absent because `gig-up.sh` — which regenerates it, and
+   whose checker exits non-zero correctly — had never been run on this box.
+6. **Freshness checks compare what they were told to compare.** Bank NAMES
+   matched while COUNTS had drifted, and the checker printed `ok` while
+   SuperDirt's own boot printed `expected 77 files, got 72`.
+7. **Ask why two numbers differ instead of assuming one rounds the other.**
+   111 vs 124 banks was a case-blind regex silently inventing two bank names
+   that do not exist, noticed only on the third sighting.
+8. **A valid measurement can be silent.** Mute sits downstream of all DSP, so a
+   muted sink still clocks the graph at the quantum — but a null sink is driven
+   by a software timer and would answer far too optimistically. The constraint
+   ("friend sleeping") forced a better test design, not a worse one.
+9. **Test the detector in BOTH directions.** Doing so caught a `WATCH` list
+   naming the process `scsynth` instead of the node `SuperCollider`, a mute
+   guard reading an error as a negative, and `re` never having been imported in
+   `rig-doctor.py`.
+
+**Deps.** The quantum question is still open and now unblocked: last night's
+`W/Q 1.30` was disk stalls, the cache is warm, and `latency-lens.py` implements
+the documented method. It needs a muted sink (or explicit consent to be audible)
+and, before any venue claim, a re-run with the UMC202HD on USB.
+
+---
+
 ## n/a — the box plays alone, and the guard that said so was lying
 
 **Description.** Answer "are we ready to take off with only this laptop?" —