Benches: keep scheduled compile + interactive_slo run; track SLOs and bench coverage gaps

[jfolcini]

Benches stay on the scheduled workflow (not per-PR) — confirmed acceptable. This issue tracks the ongoing bench hygiene that comes with that.

Current state

• .github/workflows/scheduled-deep-checks.yml runs, on schedule:
  • cargo bench --no-run — compiles all bench targets (catches bench bit-rot).
  • cargo bench --bench interactive_slo — the interactive-SLO bench.
• 26 bench targets exist under src-tauri/benches/ (agenda, backlink_query, commands, history, pagination, property, tag_query, undo_redo, loro_vs_sql_reads, interactive_slo, …).
• No per-PR bench gate (deliberate) and no automated SLO-regression tracking yet.

To do (ongoing)

1. Track SLOs over time. Capture the interactive_slo results from each scheduled run (step-summary or an artifact) and watch for regressions — flag when a tracked operation crosses its budget. (Measure-not-imagine: use observed values, not invented budgets.)
2. Coverage review — which commands need benches? Audit the command surface (the agaric_commands! set) against the existing bench targets and identify hot/user-facing commands with no bench (e.g. anything on the editor/block hot path that isn't covered by commands_bench/interactive_slo). Add benches where a perf regression would be user-visible.
3. Which existing benches need improvement? Review the 26 targets for: stale/unrepresentative workloads, missing realistic data sizes, benches that no longer compile-only-but-never-run meaningfully, and duplicate/overlapping coverage. Tighten or retire as needed.
4. Keep the scheduled compile gate green (bench bit-rot is the cheapest thing to catch).

Related: #113 (bench-gated DB perf items — uses benches to gate refactors; distinct from this hygiene issue).

[jfolcini]

Status update — first hygiene pass shipped (#1049); ongoing items remain, keeping open

This is an ongoing bench-hygiene plan, so it stays open by design. Progress against current main:

Done (PR #1049):

• SLO capture over time — scheduled-deep-checks.yml:126,141 tees cargo bench --bench interactive_slo into a GITHUB_STEP_SUMMARY table (observed vs the bench's own budget) and uploads the log + criterion estimates as a 90-day artifact.
• Coverage gap (partial) — added benches/page_load_bench.rs (load_page_subtree, get_blocks), two previously-uncovered hot read commands.
• Scheduled compile gate (cargo bench --no-run) stays green.

Remaining (ongoing):

1. Over-time regression diff/alerting — current step-summary is point-in-time; no automated cross-run regression flag yet.
2. Command-surface coverage — still-uncovered hot commands noted in ci(bench): capture interactive_slo SLOs + add page-load benches (#978) #1049: delete / todo-batch / add-tag, list-pages-metadata, find-undo-group.
3. Existing-bench review — audit the ~26 targets for stale/unrepresentative workloads, missing realistic data sizes, and duplicate coverage; tighten or retire.

Keep open as the living hygiene tracker (distinct from #113, which gates DB refactors).