[swerner]

One assistant response can contain two edits to the same file. Add deterministic regression coverage proving that both successful edits survive Pebble's coding-agent dispatch, including profile aliases and permission middleware. A test-only scheduling override reproduces the original failure: both tools report success, but the last write loses the first edit.

Inspired by @denkhaus's lost-edit report and delayed-write regression in fabro-sh/fabro#787. The tests adapt that reproduction to Pebble's existing batch scheduling; this PR adds no production fix.

The six new tests cover same-file edits through Anthropic, Claude 5, and Kimi profiles; the negative control; different-file scheduling; overlapping apply_patch calls; ordered whole-file replacements; and Kimi append calls. They use scripted provider responses, mutable state, and paused-time delayed writes, with no paid model calls or two-reader barrier.

The assertions preserve the current scope: a mutation serializes the entire batch, including different-file calls; whole-file replacements remain last-write-wins; separate sessions and workflow stages are outside this guarantee.

Validation:

• The new tests and existing tool_scheduling, tool_registration, and e2e suites passed on both Pebble eb3391c and Fabro's pinned Pebble e90513f: 20 tests on each revision.
• The pinned verification used Fabro's runtime dependency versions with unchanged Pebble Rust source; only the snapshot-test dependency differed.
• All routine verification components passed: pinned formatting, workspace Clippy, documentation tests and builds, feature checks, workflow audit, whitespace checks, and the complete Nextest suite (1,677 passed; 1 skipped). The initial sandboxed suite could not bind its fake HTTP servers; rerunning mise run test with local socket access passed.