scene_close discloses discarded unsaved edits (best-effort, 4.7-only)

scene.close closes an open scene tab. Until now it closed silently, giving the caller no signal that the tab held unsaved edits — closing throws those away, and an agent driving the tool had no way to know a discard happened. For a 1.0 that is honest about destructive actions, the tool must both declare itself destructive and, where the editor can tell us, disclose whether a discard actually occurred.

The blocker is a version constraint. A reliable “is this scene dirty?” query is only exposed to GDScript in Godot 4.7: EditorInterface.get_unsaved_scenes() (returns the PackedStringArray of unsaved scene paths) is ClassDB-bound in 4.7 and absent in 4.5/4.6. On 4.5/4.6 there is no bound dirty query at all — EditorUndoRedoManager.is_history_unsaved is unbound, and only the mark_scene_as_unsaved setter exists — so the editor cannot answer “was this tab dirty?” before it closes.

Decision

Declare destructive on every version; disclose the discard best-effort, 4.7-only.

• destructiveHint: true + honest description on all versions. The server tool metadata now marks scene_close destructive and the description states that unsaved edits in the closed tab are discarded (save with editor_save_scene first). This is version-independent — it is metadata and prose, not a runtime probe.
• unsaved_changes_discarded: bool in the response — only where detectable (4.7+). Before the tab is switched/closed, the toolkit reads get_unsaved_scenes() (gated by EditorInterface.has_method("get_unsaved_scenes"), the codebase feature-detect idiom, not a version-string compare) and reports whether the closed path was in that set. Below 4.7 the field is omitted entirely — absence signals “not determinable here”, deliberately distinct from a present false, which would falsely claim a detected-clean tab.
• The tool always closes on every version. Only disclosure fidelity differs by version; behavior (the tab closes) does not.

Deliberately NO refuse-to-close (force) guard

A guard that refuses to close a dirty tab unless the caller passes force: true was considered and rejected for 1.0. Such a guard needs only-when-dirty detection — which is 4.7-only — so any guard is necessarily either version-divergent (refuses on 4.7, silently closes on 4.5/4.6, the worst kind of inconsistency for an agent-facing tool) or fires on every close regardless of dirtiness (forcing the flag on saves that never had edits, training callers to pass force reflexively). Neither is acceptable.

The cross-version path to a real guard is a content-hash comparison (hash the on-disk scene vs the in-editor state to detect dirtiness without a bound query), which works on 4.5/4.6 too but is a larger design. It is deferred and filed at Plan/Ideas/PostRelease/2026-07-08-scene-close-dirty-guard-hash.md. Recorded here so a future contributor does not re-litigate the “just add a force flag” shortcut without seeing the version-divergence trap.

Consequences

• The tool is honest by default: destructive metadata + a save-first description on every supported version, and a concrete per-call discard signal on 4.7.
• Callers must not rely on unsaved_changes_discarded being present — it is a 4.7+ affordance. On 4.5/4.6 they should treat the close as potentially discarding and save first if that matters (the description says so).
• A stricter refuse-to-close guard remains open, gated on the deferred hash-based detection, not on the 4.7-only query.

Iteration: Plan/ExecutionPlan/archive/41o-nonies-pre-1.0-tool-polish.md; grill log Plan/Reference/GrillingSessions/2026-07-08-41o-nonies-tool-polish.md.