PLN, after a change of mine reached his ears before a check did: "we gotta make
things work bro we cant break as we move i thought we had clarity and confidence by
now :) go on, take the longshorter path not hte immediate one".

Right, and the honest diagnosis is not "be more careful". It is that this toolbox
had a whole missing category of check.

Everything in it is COLD:
    pvlint            parses .tidal text
    silent-eval       queries patterns against an EMPTY control map, in a fresh ghci
    surface-columns   reads .tidal text
    check-boot        loads BootTidal in a throwaway ghci
    check-mix         reads the SAVED Ardour session

All five were green all afternoon. Meanwhile gMask sat ARMED AT 127 on the live
board from 15:36 to 16:50 — a global chopping the last eighth out of every bar of
every orbit wrapped in gM1/gM2/gM3, which is every orbit of every track. Nothing
was broken. A switch was on, and NOT ONE tool in the chain looks at the switches.
PLN found it by ear, and the ear should never be the smoke detector.

`gig-log preflight` closes that. It reads the live recorder's surface state and
returns a verdict on the controls that can silence a whole rig from outside any
pattern: gPanic armed, any gMute engaged, a DJ filter parked in a band-killing
position, gMask armed if it is still active.

== THREE DESIGN DECISIONS, EACH FROM A PAST FAILURE ==

1. A STALE LOG FAILS. Exit 2, and distinct from exit 1 (surface unsafe) so a caller
   can tell "I could not check" from "I checked and it is bad". A preflight that
   reports SAFE because it read yesterday's session converts an unknown into a
   false reassurance, which is precisely the shape of every bad afternoon this rig
   has had. is_live() already existed; this is the first thing to gate on it.

2. gMASK'S MEANING IS READ FROM BootTidal.hs, NOT HARDCODED. ^41 was a global gate
   this morning; since 37857225 it is d1's gate and gMask is `id`. Same CC, opposite
   verdict. Hardcoding either answer would be wrong half the time and would go
   stale exactly the way every other cached binding here has
   (feedback_stale_binding_pattern). An unreadable BootTidal assumes the mask is
   STILL ACTIVE — the cautious read, because guessing "retired" silences a real
   FAIL.

3. IT NEVER FAILS ON THE STATE BUTTONS. Gates on 41-44/57-60/76/89-92 are gestures
   PLN arms on purpose. They are LISTED ("armed on purpose (not a failure)") so
   nothing is silently on, but they do not block. A gate that cries wolf about
   intentional performance state is a gate that gets ignored, and then it is worth
   nothing on the night it matters.

It also names what it cannot see rather than implying completeness: Ardour faders
are invisible over MIDI, so the output points at check-mix.py and says that reads
the SAVED session.

== AND IT IS WIRED IN, WHICH IS THE ACTUAL POINT ==
gig-up.sh now ends with a readiness gate that runs BOTH preflight and check-mix.
Everything above that line only proves processes STARTED; these two are the only
steps that ask whether the rig will make a SOUND. Neither aborts the launch — you
may be starting up precisely to fix them. This is the "real readiness gates" half
of feedback_simple_setup, and it means the surface check happens whether or not
anyone remembers to ask for it.

Verified: live run reports SAFE with the board as PLN just left it (48 controls
touched, gMask correctly detected as retired, ^93 untouched and flagged as sitting
at the #55 seed). Negative-tested against a truncated log: exit 2, "recorder is
NOT running", no verdict offered. 5 new tests (309 total), including the
load-bearing one that a stale log must FAIL rather than pass, and that a
COMMENTED-OUT `gMask = id` does not count as retired.