mind: the compile-axis arc — re-scoped, issued, shipped, recorded

[Jammy2211]

The Mind side of the compile-axis arc, from re-scope to completion records. All
four code PRs are already merged; this is the ledger catching up.

Repo	Commit	What
PyAutoBrain	cd79005 (#219)	campaign --axis compile
autolens_profiling	355d555 (#104)	cache_state, pins, dashboard, + pytest in CI
PyAutoBrain	74d7b1b (#220)	ingest --axis compile
PyAutoBrain	a50efc3 (#222)	triage --axis compile, arc closed

The re-scope

draft/feature/profiling/profiling_agent_jax_compile_time_scope.md (2026-07-14) asked
the Profiling Agent to measure compile time of the release-validation heavy scripts so
we could speed them up. Both halves of that framing were spent:

• The speed-up shipped. The Split Mind prompt-file lifecycle into draft/ active/ complete/ #71→lifecycle.py record: paired dated record for new completions (#71) #74→arxiv digest: author bylines + longer highlight summaries #77 arc landed the persistent compilation cache
  (117.0 s → 2.3 s CPU; 5517.8 s → 937.1 s A100) and --xla_gpu_autotune_level=0 (17×
  on the pathological cold probe), and ruled out source jit-boundaries by evidence.
• The release-scripts leg belongs to another conductor. Script-suite cost is
  hygiene's perf mode by both agents' documented boundaries — it had been moved out of
  profiling's staged future modes once already.

What survived is regression surveillance: both wins are settings, so a config
drift or an XLA_FLAGS clobber (the PyAutoNerves#127 failure that went undetected for
two months) puts the 70-minute worst case back with nothing failing. Split into three
sequential phases; the workspace-side jax_compile_time_profiling.md was absorbed into
phase 2, since the dashboard and the agent reading it are two halves of one thing.

What this PR contains

• the re-scoped parent prompt (original text preserved under a <details> for provenance)
• three completion records under complete/2026/08/
• complete/index.md regenerated (966 records)
• jax_compile_time_profiling.md marked absorbed

active.md is back to just the two release-drive entries. The three phase prompts moved
draft/ → active/ → complete/ through the lifecycle rather than jumping states.

The traps the records preserve

Every one surfaced by running the tools against real data, not from the design:

• cache_dir is non-empty on COLD rows too — the cold run is the one that populates
  the cache, so warmness cannot be read from it or from the tag (~40 ad-hoc spellings).
• hardware alone pooled two machines — one local_cpu label spanned a laptop (66
  records) and a 32-core RAL node (12). The first rendered dashboard put pins from both
  side by side under one heading. hostname joined the comparability key.
• Pins auto-following the newest row would have baked a regression in permanently,
  reporting all-clear forever — exactly inverted from the tool's purpose.
• Rows predating a pin are history, not drift — the first ingest run flagged four
  measurements the pin had been chosen over.
• export_probe.py / trace_profile.py share the results tree with a different
  schema, so their 4 records are sibling-instrument output, not corruption.

Two corrections to my own earlier claims, recorded rather than quietly fixed

• The sweep grid is 21 cells, not the 24 first written in the re-scope; only
  imaging/mge/hst and imaging/pixelization/hst are on-grid.
• The workspace test that looked red outside CI was a missing matplotlib
  (aggregate.py imports it at module scope), not a defect. It passes in CI — now
  provably, since lint.yml gained the pytest step it never had.

Verification

lifecycle.py check: OK · 127 Mind tests pass · all three modes verified end-to-end on
merged main against the merged workspace (campaign 11/147, ingest 25 pins / 0 drifted /
0 unpinned, triage 0 findings).

---

Generated by Claude Code