fix(studio): rebase the debounced write on the current document

[LeadcodeDev]

Severity Medium, category correctness. Location: crates/rustmotion-studio/src/editor/inspector.rs:2119

Impact

write_prop/write_root_field/write_content capture m.raw (the ENTIRE document) at edit time, then 250 ms later serialize that snapshot over the whole file. If the notify watcher reloads the model from an agent/editor write during that window, the payload's raw is already stale and the flush wipes the agent's change completely — not just the edited pointer. write_and_note then registers the result in the self-write ledger, so app/mod.rs:141 makes the watcher skip the resulting event: memory holds the agent's version, disk holds the studio's stale version, and nothing ever reconciles them. No error is shown (m.write_error = None on success). This is the crate's headline workflow (baseline/diff review of agent edits), so the race is routine, not exotic.

Fix

Before flushing, re-read the file and verify it still matches what the payload was derived from (hash the source at snapshot time, compare at write time); on mismatch, re-apply the single pointer mutation to the CURRENT disk content instead of replaying the stale whole-document snapshot, or surface a conflict in write_error. Do not register a self-write for a flush whose base no longer matches the disk.

Evidence the audit read

let task = spawn(async move {
    tokio::time::sleep(std::time::Duration::from_millis(250)).await;
    *debounce_slot.borrow_mut() = None;
    let snapshot = std::fs::read_to_string(payload.path()).ok();
    let result = perform_write(&payload);
// ... perform_write, JSON branch:
} else if let Some(updated) = set_field_value(raw.clone(), pointer, field, value.clone()) {
    let s = serde_json::to_string_pretty(&updated).map_err(|e| format!("json: {e}"))?;
    write_and_note(path, &s)?;

Based directly on the chantier branch.

Part of the September 2026 audit remediation chantier. Refs #220 (RM-13).