Timeline plugin: date selector

[sandesh-sp]

Replaces the date selector placeholder with the real control: the current date in the active time mode's format, opening a popover picker whose fields match that mode — year stepper, month grid, day calendar, or datetime field. Picks are validated and clamped to the timeline range; out-of-range months and days are disabled.

Stack — merge in order, starting from the base:

1. Timeline plugin: shell and component placeholders #239 shell and component placeholders → development
2. Timeline plugin: timeline view #240 timeline view
3. Timeline plugin: date selector #241 date selector
4. Timeline plugin: playback controls #242 playback controls
5. Timeline plugin: time mode control #243 time mode control

[sandesh-sp]

Review pass over the stack (#239–#243). Seven issues in this PR, pushed as 0cd3c48. Rebased on the fixes in #239 and #240.

The big one: local time vs UTC

MMGIS core is UTC end to end — TimeControl.globalTimeFormat = utcFormat(...), times stored as toISOString() Z-strings, TimeUI on moment.utc throughout. This selector was entirely browser-local: a grep for moment.utc|utcFormat|getUTC across the plugin returned zero hits.

Two consequences, with currentTime = 2025-03-10T00:00:00Z:

• At America/New_York the header rendered Mar 9 while core's own TimeUI showed 2025-03-10 on the same screen.
• Clicking "10" in the day grid emitted 2025-03-10T05:00:00.000Z, five hours off the day boundary the data is keyed to. At UTC+10 the same click emitted 2025-03-09T14:00:00.000Z — time-enabled tile URLs built for the wrong day.

All 32 date call sites across DateSelector and DayCalendar now go through moment.utc. Everything else about the date handling was already careful — every path into onDateChange goes through commit, which validates isValid() and hard-clamps into range, so no invalid or out-of-range Date can reach toISOString(); the day grid walks real dates rather than adding 24h, so leap years and DST are correct. Getting all that right and then reading the viewer's zone was the one gap, and the fix was mechanical.

Also fixed here

• Stale date half in HOUR mode. inputValue was seeded only on popover open and never re-synced, and the HOUR branch read it in preference to the prop. Open the picker at 2025-03-10T20:00, let playback cross midnight, then edit the time field → it wrote the stale date back and rewound the map a full day. The date half now derives from selectedDate; the inputValue state is gone.
• Two-digit months were unreachable. A [viewMonth] effect rewrote monthText on the same keystroke handleMonthInput accepted: typing "1" for "10" strict-parsed to January and the field was immediately rewritten to "Jan", so the next keystroke gave "Jan0". October through December could not be entered by number, contradicting the component's own comment. Text is now rewritten only when an arrow or a snapped year moves the view.
• Popover had no dialog semantics. It portals to the end of document.body with no role, no Escape, no focus management — so opening it left focus on the trigger and Tab walked the entire MMGIS UI before reaching the calendar, with no way to dismiss in place (Enter can't submit either: 2+ text inputs block implicit submission). Now role="dialog", Escape-to-close, focus moved in on open and restored on close, and aria-haspopup/aria-expanded/aria-controls on the trigger.
• Dangling htmlFor. <label htmlFor="date-input"> rendered in all four modes but that id only exists on the year input, which renders in YEAR and MONTH. Replaced with a title element and aria-labelledby on the form.
• Day cells now carry a full-date aria-label instead of announcing a bare day number with no month or year.
• Calendar focus indicator was folded into the hover rule: outline: none plus a #e6f4f6 tint at ~1.15:1 on white, indistinguishable from hover — and :focus-visible (0,2,0) outranked --selected (0,1,0), so focusing the selected day stripped its fill and the grid looked unselected. Split out with a real ring; the ringless reset now applies to pointer focus only.

tsc --noEmit clean, 898 unit tests pass.

One thing left alone: the "Compare date" button still has no onClick. Looked like intended future work rather than a defect to fix here — say the word if it should be wired or removed.