[cat0825]

Four things the tray needed, all in the same subsystem. If you would rather take them one at a time, the two bug fixes (Tcl discovery, the refresh button) are the ones worth splitting out; the rest is presentation.

1. The icon was not authored for a menu bar slot

A slot is 18pt, so 64px covers it at 3.5× and the Windows tray at 4×. The artwork was a 1254px canvas whose ring interior was filled with opaque white — invisible against a white page, a white blob on a dark menu bar — and 383,807 bytes of it.

Rebuilt from the same artwork:

• interior and background made transparent, so only the ring remains
• cropped to the ink and resampled to the 64px that ICON_PX already hands over
• menubar-icon.png for a light menu bar, menubar-icon-dark.png for a dark one, where the navy arc is lifted toward a light slate and the blue accent is left alone. custom_icon_path() already prefers a -dark sibling, so this replaces the per-pixel lighten_for_dark() pass with a hand-tuned file

383,807 bytes → 8,789, and the drawn fallback is untouched — it still renders at 64px with a transparent corner and a white centre, exactly as test_tray.py asserts.

A precedence bug that shipping an icon exposes

custom_icon_path() returns the first file that exists, and the bundled default was listed before ~/.headroom/icon.png — so once an icon ships, a per-user icon can never win. The per-user file now comes first, and a test pins the order. This was dead code before; it would have become a silent bug the moment an icon landed.

2. A detached venv process lost Tcl

A venv interpreter resolves tcl_library from its base prefix, but a process started with start_new_session loses that and dies with:

Cannot find a usable init.tcl

That reads as a missing Tk, but it is a search-path problem — and it is exactly the setup the README recommends (~/.headroom/venv plus requirements-desktop.txt), so the macOS menu bar could not start at all in the documented configuration. Setting TCL_LIBRARY/TK_LIBRARY at import time makes every launch path behave the same.

3. The refresh button could wedge permanently

The reader thread published its result with a plain put(), so busy was cleared only when a result arrived. A reader that raised outside the caught set — or hit a full queue — left every later refresh a silent no-op. That is exactly what "clicking refresh does nothing" looks like from the outside. The result is now published from a finally, a full queue is drained rather than dropped, and a click shows the loading label immediately because the percentage rarely changes between two reads.

4. The card itself

• The refresh button is a flat pill rather than a tk.Button — Aqua ignores a button's background, so the old one rendered as the system default, not as designed.
• The meter is a thin rounded line instead of a heavy bar; the percentage is split from its label; the mood caption sits under a hairline.
• Fonts come from one UI_FONT/CJK_FONT pair, so the card follows the platform instead of hard-coded Segoe UI/Microsoft YaHei UI.
• The mood image is now clickable in the card as it is in the web dashboard: it hops and plays one of the bundled clips with afplay, rotating so repeated clicks do not repeat the same sound. Tk cannot decode video, so the tray card stays audio-only.

Docs

The icon was undocumented — grep HEADROOM_ICON README.md returned nothing. README, README.zh-CN and SKILL.md now describe the bundled pair, the -dark sibling, HEADROOM_ICON, ~/.headroom/icon.png, and the drawn fallback.

Tests

TrayIconTests (custom icon resolution, the -dark sibling, the per-user precedence), TclLibraryTests (the search path, and that existing variables are left alone) and two new desktop tests for the reader and the mood click. Full suite passes on CPython 3.10, 3.11, 3.12 and 3.13, and with the tray dependencies installed so the tray tests actually execute rather than skip.

One note on that last point: test_tray.py only runs on Windows, so an earlier revision of this branch that raised ICON_PX to 72 passed locally and failed there. Reverted — 64 is the size draw_icon is drawn on, and raising it without rescaling that geometry would have left the drawn fallback off-centre.