perf(log-viewer): 24% faster Call Tree grids, and a highlight that follows your row

[lukecotter]

📝 PR Overview

A merged call-tree row was identified by joining its bucket keys into a string, which was most of the cost of the grid builds. Separately, a bottom-up caller row reported the calls it counts rather than the frames it is, so every caller depth highlighted the same leaf frames however deep you picked.

The grids now group on the interned key ids the inspector already uses, and a row now reports the frames it stands for, so stepping down the callers walks the highlight up the stack. Three navigation bugs found while walking that change are fixed here too.

🛠️ Changes made

• Grid builds on interned key ids — both builders take the log's KeyPathIds and group on integers instead of joined key strings. On a 95MB log (864,216 lines, 431,307 calls) toBottomUpTree goes 497ms → 377ms and toAggregatedCallTree 337ms → 308ms. Medians of four runs; run-to-run variance on this log is about 50ms, so read the second figure as directional.
• A frame outside the log's own index is keyed but not cached — keyIdOf wrote past the end of its Int32Array, which lands as an ordinary property, so every frame built rather than parsed read back the first one's id.
• The Analysis reveal reads no occurrence — it found its bucket by scanning the 431,307 occurrences the root buckets hold between them, then listed every active row again to test one filter. One key compare over the top-level rows and one boolean read now do it.
• A bottom-up row highlights its own frames — frameEventIndexes climbs the row's own path depth from each counted call and dedupes, so the flame chart and call tree point at the frames the row is. Details still describes the calls it counts, which ride on a field of their own.
• The keyboard survives a tree-control click — the control is not focusable, so clicking it dropped focus, and the key bindings only answer while the table body holds it: the arrows scrolled the table instead of moving through it. Focus now returns on every pointer expand and collapse, not only the first one before any row is selected.
• A bucket descent waits for the render it needs — the descent skipped the wait entirely for any row already open, so it read empty children and fell back to that row: picking a deep inspector row landed on one of its callers and needed a second click. It now waits for any row whose children have not arrived.
• Analysis moves to a picked inspector row — it marked the bucket but never scrolled to it, because only a finding click revealed. It now reveals on a sticky locate and then marks, as the Call Tree does.

🧩 Type of change (check all applicable)

• 🐛 Bug fix - something not working as expected
• ✨ New feature – adds new functionality
• ♻️ Refactor - internal changes with no user impact
• ⚡ Performance Improvement
• 📝 Documentation - README or documentation site changes
• 🔧 Chore - dev tooling, CI, config
• 💥 Breaking change

📷 Screenshots / gifs / video [optional]

N/A. Nothing new appears on screen: what changes is build timing and which frames light up.

🔗 Related Issues

None.

✅ Tests added?

• 👍 yes
• 🙅 no, not needed
• 🙋 no, I need help

📚 Docs updated?

• 🔖 README.md
• 🔖 CHANGELOG.md
• 📖 help site
• 🧪 Marked any pre-release-only features
• 🙅 not needed

Two entries, for the keyboard and navigation fixes. The highlight and Analysis fixes correct the Inspector, which is unreleased, so they belong to its existing entry rather than a new one.

Anything else we need to know? [optional]

Where to start. log-viewer/src/core/log/keyPathIds.ts carries the vocabulary; read its class doc first. The two id spaces matter: the aggregated build composes outermost-first and the bottom-up build innermost-first, so ids from the two directions must never be compared.

Two questions, two accessors. locatableEventIndexes is the calls a row counts, which its totals describe. frameEventIndexes is the frames the row is, which a highlight points at. They differ only for a bottom-up caller row.

Test plan.

• pnpm lint and pnpm test.
• Call Tree, Bottom Up: expand a method to its callers. Totals and call counts read as before. Hover depths 2, 3 and 4: the lit frames walk up the stack a level at a time, and Called by names the row hovered.
• Click a row's expand arrow, then press the up and down arrows: the selection moves and the table does not scroll on its own.
• Aggregated: pick a deep row from the inspector's call tree. It lands on that row first time, not on one of its callers.
• Analysis: pick a row in the inspector's call tree. The grid scrolls to the bucket and selects it. Picking a bucket the Show Details filter hides turns that filter off, as a finding click already did.
• Both themes, and the Inspector docked at the side and at the bottom.

Known unrelated failure locally. lana/src/services/__tests__/servicesRuntime.test.ts cannot resolve effect in a worktree that has not been installed since #951. No lana file is touched here.

Reverted during review. A first attempt routed that wait through RowNavigation's pending-render flag. Tabulator dispatches renderStarted and renderComplete in one synchronous call, so the flag always read false where the wait was awaited and the indirection bought nothing, while its one live branch could only be entered by a renderStarted whose renderComplete never came — which one throwing subscriber causes, since _dispatch has no try/catch and five modules subscribe. The wait would then never settle. The inline wait is back.

Follow-ups, not in this PR.

• The reverse direction still marks on the old rule: hovering a frame marks the rows whose leaf it is, rather than the rows whose own frame it is. Doing it properly needs a walk of that frame's subtree.
• The reveal-then-mark handler is now shared in shape by four views but copied in two. DatabaseView and ApexLogTimeline mark without revealing, so "does a pick move this view" deserves to be an argument rather than a property of which handler was copied last.