fix: build_readme.py — probe correct JIT-time keys + handle NaN (Phase 4 follow-up)

[Jammy2211]

Summary

Phase 4 follow-up bug surfaced when running the dashboard generator against the first real CPU artifacts produced today on this machine. Every cell in every auto-table rendered as — even with fresh artifacts present.

Two related bugs

Bug 1: key-probe mismatch

Shipped likelihood scripts emit full_pipeline_single_jit (no _s suffix); the datacube script emits full_pipeline_cube_single_jit. But _likelihood_headline_seconds was only probing for _s-suffixed variants nested under summary / aggregate. Result: every cell rendered —.

Fix: prepend (\"full_pipeline_single_jit\",) and (\"full_pipeline_cube_single_jit\",) to the probe-path list. Older _s-suffixed and nested variants stay in the list for forward-compat.

Bug 2: NaN renders as nan s

datacube/delaunay.py emits full_pipeline_cube_single_jit = NaN when the JIT-compiled cube pipeline doesn't converge — a real signal worth preserving. Python's json.loads parses that as float('nan'), which slipped through unchanged and rendered as nan s in the table.

Fix: math.isnan guard returns None, so _format_time renders the cell as — (matches the "no data" semantic while preserving the signal that the artifact exists but the measurement didn't land).

Verification

Ran 4 known-clean scripts in the canonical autolens_profiling repo today (Intel i9-10885H, CPU-only JAX, PyAutoLens 2026.5.14.2). With this fix, the headline auto-table renders:

Section	Script	Instrument	Latest single-JIT per-call
likelihood/imaging	mge.py	hst	106.8 ms
likelihood/interferometer	mge.py	sma	38.5 ms
likelihood/datacube	delaunay.py	hannah	— (NaN: cube JIT didn't converge)
simulators/imaging.py	—	—	3.12 s total wall

build_readme.py --check exits 0 on a second run — idempotence preserved.

Test plan

• Run python scripts/build_readme.py against 4 fresh artifacts; verify cells populate.
• Re-run python scripts/build_readme.py --check immediately after; exit 0 (idempotent).
• Verify NaN cell renders as — not nan s (datacube hannah artifact).

Refs

• Phase 4 PR: feat: Phase 4 — instrument-framed README dashboard #10 (introduced the bug)
• Phase 4 follow-ups in top-level README: this fix wasn't on the deferred list because the bug was masked by the "no real artifacts exist" state at PR feat: Phase 4 — instrument-framed README dashboard #10 ship time. Will likely add a note in the README.

🤖 Generated with Claude Code