feat: run counts and greying out of impossible filter options (#296)

[timdegroot1996]

Fixes #296

Problem

The filter dropdowns are built once from the full run data and never recomputed, so they keep offering values that belong to runs another filter has already excluded. With two custom filter dimensions where filter_1=A only ever occurs together with filter_2=C, selecting filter_1=A still offers filter_2=D, which can only ever yield zero runs.

Root cause

setup_custom_filters_in_select_filter_buttons() (and the equivalents for Runs, Run Tags, Versions and Metadata) run once at startup over the whole runs array; filtering itself only happens when the modal closes. Nothing ever tells an option list what the other filters did.

Fix

Every option of the Runs, Run Tags, Versions, Metadata and custom filter dropdowns now shows how many runs remain if that option is the selection of its filter. Options that match nothing are greyed out — they stay visible and selectable, so no value disappears from the modal (the middle ground suggested in the issue).

• The counts follow the usual faceted-search rule: they are computed with every filter except their own applied, so picking a value never makes the other values of that same filter vanish.
• The amount filter is left out: "the most recent X runs" is not a category.
• In NOT mode an option excludes its runs, so its count is what is left over.
• Both the counts and the greying out can be turned off in Settings → Defaults.

To keep one implementation of the filter logic, the filter functions were split: apply_* functions hold the logic and take their selection as an argument, filter_* functions read the filter modal and call them, and the availability computation reuses the same apply_* functions. Parsed custom_filters and metadata strings are cached, which also removes a per-dimension re-parse from the existing filter pass.

Readability of the dropdowns

Putting counts and greyed-out rows in the dropdowns made it obvious that the panels themselves were hard to read, so a second commit fixes that.

• The panels had no surface of their own — only position, so the rows inside them painted the background. Those rows use --color-card, which is rgba(30, 41, 59, 0.9) in dark mode, so the modal rows underneath showed through an open dropdown and the text of both was mixed together. The panels now paint an opaque --color-popover-bg with a border, a radius and a shadow, and the rows inside are transparent.
• The search box scrolled away with its own option list; it is sticky now.
• Unavailable options were dimmed twice, with colour and opacity: 0.65, which also faded the checkbox and the count — the count being the one thing that explains why a row is greyed out. The opacity is gone (it is ignored on <option> elements in native select popups anyway). What is left is a dedicated --color-text-disabled on the label, a struck-through label and a faint row tint, with the checkbox and the count at full contrast.
• The counts trailed labels of arbitrary length, so they were impossible to compare. They now sit in their own right-aligned tabular-nums column, which needs them to be a sibling of the label rather than a child: the row is the flex container. Long values are ellipsised instead of pushing the count out of the row.
• --color-text-muted was darkgrey, ~2.3:1 on the light theme's white, which fails WCAG AA. It is #6c757d (~4.6:1) now. This token is used beyond the filter modal, so light-theme muted text is slightly darker everywhere; no reference screenshot moved.

Tests

• New tests/javascript/filter_availability.test.js: 14 cases, including the exact matrix from the issue, the exclude-its-own-filter rule, NOT mode, None, date range and suite path. JS suite: 314 passed.
• Two new robot tests in 06_filters.robot: the counts and greying out, and the two settings that turn them off.
• Full robot suite in Docker: 105 tests, 105 passed, 0 failed on the first pass (no rerun needed), both before and after the readability commit. Python: 403 passed.

Also in here

Open Filter Dialog now waits for the modal's opening animation to finish. Bootstrap ignores a hide() that arrives while the modal is still animating open, which leaves the modal on screen and makes every later click time out. That race was already flaky on main (the baseline run flaked on 08 Layout); this change shifted the timing enough to hit it repeatedly in 06 Filters.

🤖 Generated with Claude Code