START. The field run is LAUNCHED and running. LAUNCH_UTC=2026-09-12T10:26:16Z.

  subject: fp-driver-v9-bundle.sh 268436e507e3937fb244758c77e1df7131e3713b9369974b3e76ad3dcaf88e99 with r3 helper 63d0508b93a8d15fa2bc9940708c7a13ac4abda1f8ca9c9a5d32a63b1dcf3e8a
  evidence root: bundle-r3/fp-run/20260912T102616Z
  isolated home: bundle-r3/fp-home (did not exist before launch - the driver refuses a pre-existing one)

FRESH PREFLIGHT AT LAUNCH, all six, and then the driver's own copies of them agreeing from inside the run:
  my reads at 10:25:45Z - free 122.08 GiB; r3 census valid=YES cargo_build=0 runner_descended=0 shim_unresolved=0 analyzer=6 RUNNER_WORKER=ABSENT; both GitHub runners online and not busy, queued=0 in_progress=0; lane worktree clean at 53d625cd with blob c28874ef; fp-home absent.
  the driver's own records, from findings.txt:
    ELEVATION self=False (expect False)
    INSTRUMENTS hashed into INSTRUMENTS.sha256 (5 files)
    CAPACITY ok before 'preflight': free=122.08 GiB floor=96 hard_stop=32
    CI_AXIS ok at 'preflight': no non-terminal runs on BigscreenVR/spt-bs-core
    BOX_GATE ok at 'preflight': cargo_build=0 runner_descended=0 shim_unresolved=0; analyzer=6 RECORDED AND NOT GATED
    SUBJECT sha=53d625cd0bd88a04815efdf6c8209a3096bf53e8 clean blob=c28874ef6823a7bb522168514ac81c16ddd427d5
  INSTRUMENTS.sha256 inside the evidence is itself the provenance that r3 is what ran, since the driver hashes what it finds at $BIN rather than trusting placement.

SUPERVISION THRESHOLDS I AM HOLDING, against a measured baseline rather than a remembered one: baseline at launch was free 122.08 GiB, used 1739.94 GiB. I stop at free <= 32 GiB, and at 64 GiB of growth, which from that baseline means free dropping below 58.08 GiB. The driver independently enforces the 96 GiB producer floor at every capacity_gate, so a footprint that eats the headroom refuses at the NEXT producer rather than silently continuing. CARGO_BUILD_JOBS=2 is exported by the driver at :678. Cold-build duration and footprint remain unknown and I am not estimating either.

YOUR LAST INSTRUCTION IS TAKEN: liam will receive the post-build provenance ALONGSIDE the request path and nonce, and the separate stdout/stderr captures are preserved - the request file already declares redirect_stdout_to and redirect_stderr_to as distinct paths, and v9's composed run_exactly writes to exactly those two, so the two streams never merge. I will not inline the command text.

NOW IN FLIGHT: pool claim and the cold release build (cargo-class processes are up). No competing build or test execution - hertz has been told the window is exclusive until I report END, and asked to message me rather than kill anything if they need the box.

Next report is either a failure or provenance-and-phase readiness after the build, then END with cleanup evidence and pool release. I am supervising to cleanup and END, and I will not blind-retry anything.