fix(init): opt-out for project-local editor files - #1836

Thanks @sloemo01. The minimal-edit approach is the right answer to the reformat half of #1603. Re-serialising a tracked, repo-shared file end to end so that one contributor's repowise init shows up in everyone's diff as a whole-file rewrite is a real cost, and merge_json_object_member addresses it at the layer where it happens rather than papering over it downstream.

Careful detail worth calling out: returning UNCHANGED when edited == original, distinct from KEPT. Collapsing those two would have been the easy mistake and it is what makes the next point a bug rather than a catastrophe.

The caller can destroy a user's file. In claude_code.py:

action = merge_json_object_member(config_path, "mcpServers", "repowise", new_entry)
if action is FileAction.KEPT:
    action = write_json_config(
        config_path, {"mcpServers": {"repowise": new_entry}}

KEPT is the sentinel for "I cannot edit this safely": the root is not an object, mcpServers is present but not an object, or _insert_member could not find a place. In every one of those cases the fallback writes a document containing only the repowise entry. Every other MCP server the user configured and every other top-level key in .mcp.json is gone, in a file that is normally committed.

The helper's own comment says the opposite of what the caller does:

# container_key is present but not an object; leave the file alone
# rather than guessing what a rewrite would mean.
return FileAction.KEPT

It leaves the file alone, and then the caller guesses. The old code did not have this hole: it did existing["mcpServers"] = servers on the parsed document, so unrelated root keys survived even on the rewrite path.

Reaching it needs a .mcp.json with "mcpServers": null or "mcpServers": [], which is not exotic for a hand-edited or partially-migrated file. The comment says the fallback is there "rather than silently dropping the registration", but the trade as written is one missing registration against an overwritten shared config, and the second is much worse. Either keep the parsed document and merge into it as the old path did, or treat KEPT as "leave it, warn, move on".

Two smaller things:

The PR title and body say "opt-out for project-local editor files", and there is no opt-out anywhere in the diff. It is entirely the no-reformat half. #1603 has both halves and this closes it on merge, which would mark the opt-out as done when it is not. Worth retitling and changing Fixes #1603 to Refs #1603, especially since @rohitsalla has had #1133 open since 27 July doing the opt-out half in editor_integrations/. The two are complementary, not competing, and I would rather both land than have one silently close the other's issue.

Non-blocking: 270 lines of hand-written JSON position scanning is a lot of new surface in a write path. _find_object_member and _insert_member look right to me, but a fuzz-style test that round-trips a handful of awkward documents (one-line objects, trailing whitespace, nested mcpServers, a value containing }) would buy more confidence than the same number of hand-written cases.

Fix the fallback and I will take this.