feat(data_load_mcap): always-include transform-tree channels (/tf et al.) in the topic picker

[facontidavide]

Problem

The MCAP topic-picker treats every channel identically. When a user narrows their selection instead of loading everything, /tf / /tf_static are easy to leave out — and PJ4's Scene3D then silently fails to place PointCloud/Mesh3D/other 3D objects, because the transform tree never reaches the ObjectStore.

Mechanism: generic (schema, encoding) whitelist — no ROS code path

Channels whose (schema name, message encoding) matches kAlwaysIncludeRules are always loaded — pre-checked and locked (disabled row + tooltip) in the picker:

schema	encoding	convention
tf2_msgs/msg/TFMessage	cdr	ROS 2
foxglove.FrameTransform	protobuf	Foxglove

Matching is deliberately on message type, not topic name: verified with the mcap CLI on real recordings, Foxglove files publish transforms under arbitrary topic names (e.g. tf, no leading slash — example-023-av-argo-2.mcap), so a /tf name rule would silently miss them. Likewise the encoding is the channel's message_encoding (cdr), not the schema-record encoding (ros2msg) — an easy field to confuse. The loader stays format-agnostic: the whitelist is data, the mechanism knows nothing about ROS.

Enforcement is model-layer, view is garnish

reassertAlwaysIncluded() (idempotent) re-inserts whitelisted non-empty channels into selected_topics_ after every mutation point: analyzeFile() (covers loadConfig() restore of pre-feature layouts), onSelectionChanged, and btnDeselectAll. Zero-message channels are never force-included. mcap_source.cpp is untouched — it already reads only selectedTopics(), so the guarantee holds even if the view desyncs.

Also included: selection restore switched from index-keyed setSelectedRows to text-keyed setSelectedItems, the same sort-safety pattern as #188 (the table has sortingEnabled=true; index restore desyncs after a user sort). disabled_rows stays index-keyed — cosmetic-only risk, documented in the spec.

Tests

Restored the mcap_dialog_test target deleted in #144 (McapDialog previously had zero direct coverage) — now 10 tests: whitelist hit/near-miss matrix, force-include surviving stale saved configs, Deselect All, host-reported deselection, zero-message exclusion, locked-row disabled+tooltip rendering, text-keyed selection, and fixture pinning. New fixture test_dialog_whitelist.mcap (5 channels incl. a non-conventional-name Foxglove transform and a right-schema/wrong-encoding near miss) generated via generate_verification_mcaps.py. Full plugin suite: 6/6 targets green.

Codex review

Independent Codex review verdict: ship with fixes — guarantee coverage confirmed complete; its minor test-vacuity finding is fixed here (fixture pinning). Its major finding is real but pre-existing and cross-cutting: the PJ4 host applies selected_items only to QListWidget — the QTableWidget branch in pj_dialog_host/src/widget_binding.cpp honors selected_rows only (verified on PJ4 origin/main and every branch). This affects foxglove on main and #188's ros2/webrtc identically: the model round-trips fine (table selection read-back is already text-based), but visual pre-selection isn't restored on dialog open. Follow-up: small PJ4 host patch mirroring the QListWidget selected_items apply into the QTableWidget branch — fixes all three plugins at once.

Spec and plan: docs/superpowers/specs/2026-07-06-mcap-always-include-tf-design.md, docs/superpowers/plans/2026-07-06-mcap-always-include-tf.md.

🤖 Generated with Claude Code