feat(log-viewer): make the inspector call tree readable

[lukecotter]

Stacked on #876 — its three commits are the first ones here, so review from b0b1ffb3 onward. Rebases away once #876 merges.

Call tree quality

Zero-duration rows are excluded when the tree is built. Parsing ten real logs (839,740 tree nodes) puts these bookkeeping events — HEAP_ALLOCATE 385k, STATEMENT_EXECUTE 207k, VARIABLE_ASSIGNMENT 76k, VARIABLE_SCOPE_BEGIN 15k — at 82.7% of all nodes. The Call Tree tab hides them behind Details; the inspector showed every one, and on a FINEST log they consumed the 20,000-node budget before the tree reached anything worth reading.

The inspector is a summary, so it prunes them rather than offering a toggle — read the detail rows on the Call Tree tab, which is where the full log lives. realSubtree() drops a frame post-order unless it has duration, has a kept descendant, or its type reports limits rather than time (EXCLUDED_DETAIL_TYPES). The selection itself and the path leading to it always show, even with no duration of its own. The budget is still charged per materialised node, so truncation semantics are unchanged.

Bottom-Up caller rows read Calls 0. Only the seed frame incremented callCount. Every bucket in the chain now credits the call it contributed, which is what the Call Tree tab's bottom-up already does (accumulate() runs for the seed and each ancestor).

Grouped modes lead with their ranking metric — Aggregated by Total, Bottom-Up by Self. Time Order stays chronological, since that is the mode's entire point. Headers stay sortable in all three.

Truncation is visible. ScopedCallTree.truncated was computed and never read, so a subtree cut at the budget looked complete. It now reports it, beside a note that times are relative to the selection rather than to the log.

Names say what the frame is, not what event logged it

The same ten logs say the TYPE: prefix was redundant on 99.7% of rows — METHOD_ENTRY: fflib_SObjectDomain.triggerHandler(System.Type) already reads as a method — and load-bearing on 0.2%, where the text is a bare true, a line number, or empty.

So a name is now the frame's text plus its suffix, and the type returns only when the text cannot stand alone (SYSTEM_MODE_ENTER: false). One rule in eventText.ts (eventLabel) replaces three copies of the old type !== text && !excludedTypes.has(type) test, and the per-table excludedTypes sets go away.

Three frames whose text genuinely cannot identify itself gained a suffix instead:

• CODE_UNIT_STARTED was (entrypoint), wrong for most of them — 394 of 415 occurrences are __sfdc_trigger/... units fired partway through a transaction, alongside Validation:..., VF: /apex/... and Workflow:01I.... Now (code unit), which also distinguishes a batch code unit (ffrr.ffrr.ContractPopulateBatch) from a method entry.
• ENTERING_MANAGED_PKG — text is a bare namespace token (c2g, dlrs) → (managed package).
• FLOW_START_INTERVIEW_BEGIN — text is just the flow's label → (flow).

The raw type is still available: Type is now a column in Time Order and Aggregated too (Bottom-Up/Analysis already had one), hidden by default in every view and switchable from the Columns menu, where it sorts and groups like any other column.

Also here

• fix: the Details section showed Namespace and Caller namespace even when identical, which is the common case. The second row now appears only when the two differ — packaged code entered from another namespace's trigger, the case actually worth seeing.
• fix: a frame whose text is its type rendered EXECUTION_STARTED: EXECUTION_STARTED. The fallback now short-circuits when text and type match.
• fix: SYSTEM_CONSTRUCTOR_ENTRY's suffix lacked the leading space every other suffix has, so rows read <init>()(system constructor).

Verified

• tsc -b, eslint log-viewer/src lana/src apex-log-parser, prettier --check, docusaurus build — clean
• jest --runInBand — 81 suites / 1204 tests (baseline 81 / 1192)
• New tests cover the build-time pruning (a bare heap allocation is dropped, a zero-duration statement wrapping a timed method is kept, the selection survives with no duration), caller call-counts, the scope note, both namespace cases, the three parser suffixes, and eventLabel's keep/fallback branches
• Eyeballed live in a browser harness on a 19.7MB FINEST log: rows read EXECUTION_STARTED, execute_anonymous_apex (code unit), <init>() (system constructor), SYSTEM_MODE_ENTER: false; no label doubling; Type column hidden by default; inspector Details / Call stack / Call tree render correctly in a 500px panel. The two fixes above came out of that pass.