Live-log scroll fix (route C), linter pins, basis notation, monospace + test-isolation fixes

[NCCU-Schultz-Lab]

Four milestone items plus two bugs found along the way. One commit per package
per the M-CLEAN rule.

M-LOGSCROLL route C — scroll up during a run ✅ verified live

The reported bug: the Calculate-tab log snapped back to the bottom on every
printed line, so reading earlier output mid-run was impossible.

LOGSCROLL.0 (the "is route D enough?" re-check) was done empirically in a live
Voilà session, by disabling the existing guard without a code change — stripping
the class it re-queries each frame, then enabling native anchoring:

__el.classList.remove('quantui-run-output');
__el.style.overflowAnchor = 'auto';

The view jumped to the top on every line. So ipywidgets still performs the
per-line scrollTop = 0 reset, and overflow-anchor cannot defeat it — anchoring
protects against content insertion, not an explicit scrollTop write. Route D
ruled out. A second test wrapping the Output in an outer scroll container also
jumped to the top (its subtree teardown collapses the ancestor's scrollHeight),
ruling out the cheap variant of route C too.

So the re-render had to go rather than be out-raced. New quantui/live_log.py:
one <div> created once, text nodes appended to it, never re-rendered. With a
stable node, overflow-anchor: auto holds the user's position for free and a
single at-bottom check gives follow-the-tail. The per-frame pinning guard is gone.

LiveLog mimics the widgets.Output surface the app already used
(append_stdout, clear_output, assigning .outputs), so no write path
changed — _LogCapture.write, the atomic run-header write and Clear are
untouched. That kept a structural change reviewable.

The transport, and why it is not display()

Worth reading before touching this file. The first implementation pushed each
chunk as display(Javascript(...)) into a hidden Output, and silently dropped
every streaming line — the header appeared and nothing followed.

display() inside an Output routes by parent message id: the frontend
captures iopub messages whose parent matches the one Output.__enter__ recorded.
The header survived only because it is written while the Run click's comm message
is being processed. Anything emitted from the calc thread — or from an io_loop
callback — has no message in flight, so there is no parent to route by and the
payload never leaves the kernel. Marshalling to the main thread did not help,
because the constraint is the message context, not the thread.

The old Output.append_stdout never hit this: it mutates the outputs
traitlet, which syncs over the widget comm and is parentage-independent.

So the shipped design uses the two mechanisms already proven in this app:

1. a traitlet carries the data — thread-safe, no message context;
2. a MutationObserver installed once at render carries the behaviour — the
   same way the vib camera hook installs (reflections/01 Rule 7).

The observer reads chunks from mutation records rather than re-reading DOM
state, so nothing is lost when several land in one frame; sequence numbers
discard replays and console.warn on gaps.

Known trade-off: the text lives in the DOM, not widget state, so a frontend
re-render (kernel reconnect) loses it. Python keeps the authoritative copy and
resync() repaints from it — M-RECONNECT must call resync() when restoring
a live view.

M-CLEAN CLEAN.2 — linter pins

The dev env had drifted ahead of the pins CI enforces (ruff 0.15.x / black 26.x
installed vs v0.4.7 / 24.4.2 pinned), so a bare ruff check reported 205
UP045 violations CI never saw — noise that drowns real findings.

UP045 is the Optional[X] → X | None half of UP007, which newer ruff split
into its own code. The project already ignores UP007 (requires-python is
>=3.9, target-version = "py39", mypy on 3.9, 3.9 in the CI matrix), so the
split silently un-ignored a decision already made. Added UP045 beside it rather
than converting 205 annotation sites.

• pre-commit revs → ruff v0.16.0, black 26.5.1; hook id ruff → ruff-check
  (the bare id is now a deprecated alias).
• Open floors in pyproject.toml's dev extra and local-setup/environment.yml
  replaced with bounded pins — that was the root cause; nothing prevented the
  drift. black is split on python_version because 26.x requires ≥3.10 while
  requires-python is >=3.9 (CI caught this).
• Black 26 hugs a sole multiline-string argument → two test fixtures reformat;
  content and dedent behaviour unchanged.
• UP031 newly fires (4 sites). Two plain labels → f-strings; two build
  JavaScript, where %-formatting is the right tool (an f-string would need every
  JS brace doubled) → noqa with the reason.

M-UX2 UXP2.1 — basis-set notation

6-31G* and 6-31G(d) are the same basis set written two ways, and nothing in
the UI said so — a student who learned the parenthesised form could reasonably
conclude the starred dropdown entry was a different, unavailable set.

• Card gains one short line: "Also written 6-31G(d)." Deliberately one line —
  these cards exist because the old inline notes were "a lot of word clutter"
  (FR-DESCRIPTOR-CARDS); a test guards against it growing into a paragraph.
• The basis help topic carries the full explanation: star↔parenthesis table, the
  +/++ diffuse markers and why they matter for anions, and the Dunning
  contrast — cc-pVDZ has no star because polarisation is built in, and its
  diffuse counterpart is aug-, not +.
• The alias is derived from trailing stars, so a future 6-311G** entry gets
  it for free.
• Both spellings verified against real PySCF construction (identical AO counts
  for the */(d) and **/(d,p) pairs), satisfying the roadmap's "don't show
  an alias the user might type unless it works" constraint with evidence.

Fix — the live log was rendering in a proportional font

User report: the run header's ASCII wordmark rendered as garbled letters and the
Label : value provenance rows had drifting colons.

Neither was a string-building bug — _row pads with {label:<16} and every colon
lands at column 19; the wordmark is faithful figlet. _APP_CSS's system-font rule
lists .jp-OutputArea-output, which is exactly what the log rendered into, so
it was forced into a proportional font with !important. ASCII art and padded
columns both need fixed-width cells, so they broke together. Re-asserted monospace
for the log only.

Fix — the test suite read real user settings

Two quantui gpu check tests began failing after a live GPU.8 verification left
compute.gpu_enabled = false in the developer's ~/.quantui/settings.json: the
probe short-circuited at the settings gate and never reached the import branch
those tests assert on. The tests were correct; their environment was not.

test_cli.py isolated QUANTUI_LOG_DIR but not QUANTUI_SETTINGS_PATH. Fixing
it there would have been too narrow — every test constructing QuantUIApp reads
settings, and viz.default_backend / vib_framerate_fps feed construction too.
So: a session-scoped autouse fixture, mirroring the existing _isolate_results_dir.
The suite was previously green only because that setting happened to be true.

Testing

• tests/test_live_log.py (25) — the Output contract, buffering, concurrent
  appends from 8 threads, and the traitlet transport including HTML escaping and
  sequence ordering. It also caught a latent crash: the chunk sender was named
  _send, which overrides widgets.Widget._send — the comm transport — and
  broke add_class(), which app_builders calls on this very widget.
• tests/test_basis_notation.py (24), tests/test_log_monospace.py (10).
• Full suite: 1941 passed, 17 skipped (was 1882).
• pre-commit run --all-files clean at the new pins.
• Live-verified in Voilà: scrolling up during a run now holds position;
  returning to the bottom resumes following.

Also in this branch (planning vault, no code)

M-HISTORY-HARDENING closed. HIST.2's deferred latency phase closed without
optimizing: 34 real history_load_timing records give median 58 ms / p95
202 ms, and read_pyscf_log — the stage its own deferral note flagged with a

500 ms trigger — is 0.0 ms at the median. With HIST.1–.7 all resolved, the
milestone is closed and its roadmap archived.

Not in this PR

LOGSCROLL.2 — the optional "↓ Jump to latest" affordance. Native anchoring may
already be sufficient; worth deciding after living with the new behaviour.

🤖 Generated with Claude Code