Catalogue the SingleCellExperiments a session can switch between

[david-priest]

Replaces #57, which closed when the branch it was based on was deleted after merging. The commit is unchanged, and it merges into main without conflicts.

Stacked on #56 so the diff reads clean; retarget to main once that merges.

R side only. The React dataset picker that consumes availableDatasets is not here yet.

Why

GateLabR binds to one object at launch. The GateLabR-specific Shiny interface let a user move between the SCEs already in their session, and that went missing when the React interface replaced it in 1.4.0 (launchLegacyGateLabR records the retirement).

The host contract never lost the ability — its datasets field is a list and GateLabHostDatasetPort.listDatasets() returns an array. Only the launcher and the app collapsed it to one.

What this adds

file	change
R/sce_catalogue.R	list SCEs in an environment and describe them cheaply
R/host_bridge.R	manifest carries availableDatasets beside datasets
R/launch_react.R	an activate-dataset request swaps sce_state() and resends

Listing is deliberately cheap — dimensions, assay names, and whether a workspace is already stored in metadata(), never assay data. The expensive part, the per-sample binary resources, is still registered for one object at a time and re-registered on a switch.

Two refusals worth reviewing

• Activation refuses a name that is not a SingleCellExperiment, and uses inherits = FALSE so a name in an attached package cannot masquerade as a session object — activating one would write results somewhere the user cannot see.
• Switching is refused unless the request says the workspace has been saved. The whole workspace — gates, populations, scales, compensation provenance — lives in metadata(sce), so switching without a save loses it, and R cannot verify the app saved. An explicit acknowledgement turns a silent loss into an error. If you would rather R flushed it itself, say so; that is a different and arguably better design.

Tests

test-sce-catalogue.R covers enumeration (non-SCEs excluded), catalogue contents, the object launched from an expression with no name in the environment, and the activation refusals.

⚠️ Not run — the agent does not execute R here. Please run devtools::test(filter="sce-catalogue") before merging.

🤖 Generated with Claude Code

[david-priest]

Second commit (89f918c) reworks the switch after the 2026-09-11 review found the first cut switched sce_state and nothing else:

• active_name now lives beside sce_state at app level (so it survives a reload), and every write — the generic request handler, the compensation job — goes to active_name() rather than the launch-time name. Before, B overwrote A in the global environment after a switch.
• Each object has its own dataset id (.gatelabr_dataset_id_for) and therefore its own resource names, so A's workspace and memberships cannot be written into B and A's resource URLs cannot serve B's data.
• The activate reply carries ok at the top level, where the browser's response parser reads it; the nested form was discarded and every switch timed out after 30 s.
• The manifest is built for the replacement before the state swaps, so a refused object leaves the session where it was.
• A switch is refused while a compensation Apply is running, which would otherwise commit its matrix into the newly loaded object.
• The switch prints the writeback name and the precompensation note, as the launch does.
• The catalogue marks objects the descriptor would refuse (no assays, no events, missing sample column) with the reason, and recognises a workspace saved in the older plain-JSON form.

GateLab has no picker for the catalogue yet — listDatasets() is in the host contract but nothing sends activate-dataset — so this stays R-only until that is built.

🤖 Generated with Claude Code