fix(workspace): back up before writing; make edit_file/write_file replay-safe

[LinuxDevil]

Problem

• WorkspaceCheckpoints.track() wrote the file first and recorded the backup after. A process killed between the two lost the backup; on resume the interrupted write_file re-ran and backed up the agent's own content as "before", so rewind() reported restored: [file] but put back the agent's version. For edit_file the re-run failed, so nothing was recorded and rewind silently did nothing.
• Interrupted tool calls re-run on resume (at-least-once). The built-in edit_file then threw old_string was not found because its edit was already in the file, so the model was told an applied edit failed and looped.
• Docs: sessions.md said to call resume()/send() before resolving an approval after a restart, but agent.approvals.resolve()/streamResolve() already find the session from the pause record (covered by an existing test).

Fix

• track() stores the backup before the write. A path's first write in a turn records the backup; any later record in that turn, including a re-run of the interrupted call, keeps that first backup. Rewind treats a file that still holds its backed-up content as nothing to undo.
• track() also accepts a PreparedWorkspaceWrite ({ prepare, write }). The built-in tools use it, so the backup's afterHash is known before the write, a failed edit still records nothing, and a change that leaves the file as it is records and writes nothing. The plain-function form still works: it stores the backup first, then adds a second record carrying the written content's hash.
• edit_file: when old_string is absent and a non-empty new_string is present (with or without replace_all), it returns "the edit was already applied" and writes nothing. write_file was already idempotent.
• Docs: build-a-coding-agent (new "Long turns and restarts": compaction for long turns, replay safety, session.on()/onEvent for progress during resume, approvals after a restart), durable-execution, workspace-tools, sessions (corrected the approval-after-restart text, documented that resume() doesn't stream). CHANGELOG, API reports and llms-full.txt regenerated.

Verification

• New tests: a crash between storing the backup and writing (via a store whose append throws); a crash after the write for write_file and edit_file followed by a re-run, after which rewind restores the original; the plain-function track() crash path; already-applied edit_file with and without replace_all, plus the cases that must still fail; session.on() and onEvent seeing a resumed turn's events.
• npx tsc --noEmit, npm run lint, npm run build clean. vitest run src/tools src/session src/execution scripts src/createAgentApprovals.test.ts: all pass (2011 + 102 tests).

🤖 Generated with Claude Code