Skip to content

design/88 block 88b: package analysis on ND-TIFF run_mda - #59

Merged
zacsimile merged 9 commits into
mainfrom
block-88b
Oct 18, 2026
Merged

zacsimile merged 9 commits into
mainfrom
block-88b

Conversation

@zacsimile

Copy link
Copy Markdown
Member

Block 88b of design/88 (E4): community-package analysis on a typed ND-TIFF run_mda.

What changed

  • run_mda joins ANALYSIS_TOOLS through the existing attachment path (prepare_package_analysis / AcquisitionAnalysisJob); no second path.
  • The job is bound to the datastore's real save path, from the first non-null getSavePath() the supervisor's worker observes (MM suffixes every dataset, even the first).
  • The supervisor's lifecycle drives the job: acquisition outcome, then writer completion, once each, including for Stop-cancelled and unterminated runs. A job-only finally net covers a BaseException; the MDA's own outcome is not relabelled.
  • Live analysis_metrics progress on the working line, from one helper shared with _acquire_with_hooks.
  • analysis_dispatch reported as its own phase in duration_breakdown; a run without analysis does no extra work.
  • Refused before any confirmation or hardware effect: analysis with a TIFF save mode, and any analysis_feedback key (design/89 must widen run_mda deliberately).
  • Export matches _emit_timelapse: attachment arguments do not enter the standalone script.

Evidence so far

  • Spike (design/88-spike-growing-reader.*, demo machine, 2026-10-18): <dataset>/analysis/<job>/ created inside the growing MDA dataset changed nothing — MM loadData and ndstorage reopened 1000/1000; 87a's fixture read every frame, metrics contiguous 0–999; frame-gap p50 11.28–11.32 ms observed vs 11.32–11.50 ms control (n=2, unscored).
  • Review: one product round (cancellation net moved from the acquisition to the job; test fake no longer swallowed duplicate deliveries) and one gate round (B4 now FAILs a forwarding defect rather than NOT EXERCISED; pre-block selftest reaches verify; session run lengthened to 12000 frames).
  • Coordinator full suite at 03ab1de: 5963 passed, 107 skipped.
  • Gate selftest at 5d418ec, 512 × 512 real NDTiff and both real fixture workers: 10/10 limbs PASS, 11 isolated mutations each trip only their limb. On 5a1b16f with --pre-block: exit 1, eight limbs FAIL or NOT EXERCISED.

Gate (demo machine, pending)

design/88-block88b-demo-gate.md. Cadence bound: median photon/none duration_s ≤ 1.10 over five scored repetitions (operator, 2026-10-18). The session also records, verbatim and unscored, what the agent says about Stop before a zero-interval run_mda (R01).

@zacsimile

Copy link
Copy Markdown
Member Author

Demo gate results (scored from artifacts)

Run 1 (10:20): measure 17/17 runs; session died at step 1 on the gate's own ask (87a's ask(prompt, key) won the MRO). Fixed in bc30895; the selftest now binds every session prompt.

Run 2 (10:34): 10/10 PASS, operator store restored.

  • B1: every job bound to the MM-suffixed datastore path, output under analysis/<job_id>/.
  • B2/B3: from the job records, acquisition then writer, once each, 7 completed + 2 Stop-cancelled runs; every job read every saved frame.
  • B4/S2: 25–95 package lines before acquisition end; working.png at frame 263 shows four metrics "up to frame 255".
  • B6: photon/none median duration_s 1.064, bound 1.10 (n = 5 per arm).
  • B7: TIFF and analysis_feedback refused, no confirmation, no engine start, no folder.
  • S1: session 12000-frame run Stop-cancelled at 1,726 frames; job saw cancelled then writer.
  • R01 (unscored): the agent correctly said Stop/Ctrl-C abort run_mda within about a second; it misnamed Stop as "the MMStudio MDA control".

Not attributed: both gate runs imaged 1000 frames in ~15–17 s instead of 11.3 s in both arms. The spike, rerun unchanged outside MicroClaw at 10:43, was equally slow, so the state is the machine's. The operator accepted B6 as measured; the package-load effect in that state is R157.

@zacsimile
zacsimile merged commit 860daac into main Oct 18, 2026
2 checks passed
@zacsimile
zacsimile deleted the block-88b branch October 18, 2026 09:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant