[moriyoshi]

A tackle-todos sweep that closes ten open TODO.md items in five independent units, one commit each, plus a docs commit.

Changes

• TUI (imbh-tui):
  • The waterfall is fetched once the cursor settles (120 ms debounce, Enter skips it), and the last 16 traces are kept in an LRU.
  • The waterfall pane has its own spinner.
  • Ctrl-C cancels a query in any mode and never quits; Esc cancels while a query holds the keyboard.
  • Trace ids are kept on the snapshot instead of being re-parsed from the row text.
• Projection pushdown (imbh-query): each segment's Parquet reader gets a ProjectionMask matched by column name, and the buffer is coerced to the same projected schema.
  • Guarded by an on/off batch-for-batch equality test over 35 queries on a schema-evolved corpus.
  • Measured: logs count(*) 8.1 → 1.7 ms; spans hex(trace_id) = … 7.1 → 2.3 ms; matches(name, …) 10.8 → 4.0 ms at mid cardinality.
• Equality filters (imbh): attr_eq_sql compares a promoted key without CAST, about 0.1–0.5 ms faster per 100k rows, with unchanged results.
  • The PromQL-agreement test moves into imbh-lgtm and now calls the real collapse_duplicate_samples.
• Compression codec (imbh-storage / imbh): the codec is durable in db.info, like promote and retention, so the housekeeper and compaction write with the host's codec.
  • A capped pass takes partitions oldest-first, ordered by day, then earliest segment time, then table order.
• Docker (imbh-server): a remap script's .resource now refreshes after network discovery.
  • The old seed also stripped container.network.* from records whose script set .resource.
  • New opt-in test of the plugin against a real daemon.

Behaviour changes worth a look

• Db::compact_bounded spends its budget oldest-first. Before, it worked through every logs partition before any spans partition.
• ScanStats::bytes_scanned counts only projected columns, so it reads about 0 for count(*).
• App::apply (TUI) releases the input lock only for a current-generation result.
• Every fresh snapshot clears the TUI waterfall LRU, auto-refresh included.
• The public API changes are additive only: Db::compression() and read_compression. A database written by 0.9.0 opens unchanged.

Not covered

• Nobody exercised the TUI interactively; it is covered by unit tests and TestBackend renders.
• The daemon test has not been run against a real daemon. To run it: IMBH_DOCKER_DAEMON_E2E=1 cargo test -p imbh-server --test docker_plugin_daemon -- --ignored --nocapture.
• Deferred to a 0.10 API pass: a traces().get() lower bound, time-bounding the metrics dimensions/series/exemplars APIs, and #[non_exhaustive] on DbStats.

Verification

cargo fmt --all --check, cargo build --workspace, cargo clippy --workspace --all-targets -- -D warnings and cargo test --workspace pass: 779 passed, 0 failed, 5 ignored.

🤖 Generated with Claude Code