perf: speed up filtering and graph data preparation for large datasets (#364)

[timdegroot1996]

Fixes #364

Problem

On large dashboards, applying a filter spent most of its time in our own JavaScript rather than in Chart.js. On a 504-run / 208k-test dataset, filtering took 40–66% of the time to apply a filter. #364 has the full profiling.

Changes

• filter_data: looks up run starts in a Set instead of scanning an array for every suite, test and keyword row (rows × runs).
• Timestamp transforms: the milliseconds, timezone-conversion and timezone-display copies are cached per settings combination instead of copying every row on every filter apply. The filter option availability now uses the same cache instead of its own.
• Timeline views of Most Failed, Most Flaky and Messages: rows are grouped once by label and run start (group_timeline_values) instead of scanning all rows for every run and label. With a message config, each rule's regex is built once.
• Suite section filters: get_suite_data_exclusion reads the selected suite filters from the DOM once and returns the per-row check. exclude_from_suite_data used to read them for every row.
• Smaller fixes:
  • sort_wall_clock computes its sort keys once.
  • The test select no longer does a quadratic duplicate check.
  • The base64 payload is decoded with a plain loop instead of Uint8Array.from with a callback.
• Message config timeline bug: it relied on a block-scoped function and an undeclared variable, which both throw in strict mode. They are now declared properly.

Results

504 runs / 208k tests, headless Chrome, median of 3:

Scenario	main	this PR
Apply filter, default amount	1.18 s	0.37 s
Apply filter, all runs	8.07 s	3.14 s
Initial load	3.19 s	2.29 s
Messages with a message config, all runs, 306 runs (graph update)	1.49 s	0.65 s

Chart.js is the largest remaining cost. The follow-up items are listed in #364.

Tests

• Unchanged output: I ran the same scenarios on main and this branch and compared chart data, filtered data and select options:
  • scenarios: the timestamp settings, default and all-runs amounts, bar and timeline views, suite paths, and a suite folder
  • dashboards: base, medium, and medium with a message config
  • All 1080 comparisons were identical, with no page errors.
• JS unit tests: 373 passed. New tests cover:
  • the transform cache
  • stable wall-clock sorting
  • group_timeline_values
  • get_suite_data_exclusion (the four selection cases, plus reading the DOM only once)
  • the Messages bar and timeline with a message config
• Robot tests (Docker): 124 tests, 124 passed, 0 failed. This run was before a comment-only cleanup.

Second commit: Overview, Tables, data loading

• Overview: a run-card donut is created only when its card comes near the viewport, and destroyed before the card is removed.
  • Each update used to create every donut twice and never free the old charts, so memory grew with every filter change.
  • Update with 504 runs: 4.0 s → 0.07 s.
• Tables: the columns get the type DataTables detected, so it no longer checks every cell on every update. version is left to detection. Sort and search results are identical to main. Update: 3.9 s → 2.4 s.
• Data loading: the payloads are inflated in parallel with the native DecompressionStream (based on @HuntTheSun's branch), and pako is removed. main() awaits load_data() before anything reads the data.
• Tooltips: build_tooltip_meta parses each run start once (180 → 51 ms).
• Initial load, 504-run dashboard: 3.8 s → 2.3 s.
• Browser support: DecompressionStream needs Chrome 80+, Firefox 113+ or Safari 16.4+.
• Tests: JS 373 passed, Python 406 passed, robot (Docker) 124 passed.

🤖 Generated with Claude Code