md-basics tutorials: statistical mechanics you can watch; DriveFS mount tests + truncation fix

[andeplane]

Six new tutorials (md-basics/)

Statistical mechanics made visible — each notebook takes one famous piece of physics and makes it happen live in the browser. All simulation decks were prototyped and validated in node against the real wasm first, then every notebook verified headlessly (playwright, real browser, full cell-by-cell execution):

Notebook	The physics	Validated result
01-maxwell-boltzmann	Every atom starts at the same speed; collisions assemble the M-B distribution (H-theorem live)	T stays at v₀²/2; arcsine → Gaussian
02-condensation	Quench a 2D vapor below coexistence; droplets nucleate and coarsen (colored by pe/atom)	latent heat visible in pe(t); optional KOKKOS "go big" cell
03-real-vs-ideal-gas	Measure Z = P/ρT across (ρ, T); overlay B₂(T) computed in numpy from the pair potential	virial line matches at low ρ; Boyle T ≈ 3.42
04-reversibility-and-chaos	Loschmidt's paradox: flip every velocity, un-melt a crystal; then find the chaos horizon	return error 1e-12 at 500 steps → 3.6 at 16000
05-thermal-conductivity	Müller-Plathe imposed flux + ave/chunk profile from a file → κ	κ ≈ 6 (2D LJ liquid), clean tent profile
06-polymer-chains	numpy-built FENE bead-spring chains; Rg ~ N^ν	ν ≈ 0.63 vs SAW 0.588 (finite-size, discussed in the text)

Linked from the index page (new section) and NOTEBOOK_TUTORIALS.md (Series 1 marked partially shipped).

DriveFS file-sync: full branch tests + a real bug fix

tests/drivefs-mount.spec.ts extracts the _MOUNT_DRIVEFS_JS snippet from the Python bindings and unit-tests it against the real wasm filesystem with a mocked JupyterLite ContentsAPI — 33 tests covering every reachable branch:

• toAPIPath: cwd inside /drive (prefix), exactly /drive, outside /drive, cwd() throwing, and per-call re-evaluation
• node_ops: lookup hit/miss, mknod (file + dir), getattr, setattr (mode / timestamp / size), rename, unlink, rmdir, readdir, symlink→EPERM, readlink→EINVAL
• stream_ops: open existing / missing+write (create) / missing+read (ENOENT), close with/without write flags (put vs no-put), read (normal / past-EOF), write (grow / in-place), llseek (SEEK_CUR / SEEK_END / negative→EINVAL), append mode
• mount + re-mount (unmount branch)
• LAMMPS end-to-end: write_dump, fix ave/time + close, fix ave/chunk + unfix, read_data/runFile of store-seeded files, and re-run overwrite

Bug found and fixed: truncation reaches the filesystem as a node-level setattr({size}) (Emscripten strips O_TRUNC before the stream opens), which the proxy ignored — re-opening an existing output file with "w" kept the stale tail of the previous longer content, so re-running any notebook that writes output files could corrupt them. setattr now honors attr.size (shrink / zero / extend), with regression tests for all three.

The remaining unreachable branches are defensive fallbacks (documented in the spec header); the pyodide-only Python glue (_mount_drivefs caching/fallback) is exercised by the headless notebook integration runs.

Test plan

• npx vitest run tests/drivefs-mount.spec.ts — 33/33
• npx vitest run tests/lammps.spec.ts tests/client.inject.spec.ts — no regressions
• Headless browser runs: all 6 new notebooks + index, basics/02, basics/05, materials/01 — all pass with the fixed bindings

🤖 Generated with Claude Code