- Untangle in the rewrite: `context.py` builds the final launch environment
  before `core/spawn_service.py` and `streaming_runner.py` apply Pi timeout and
  ping settings, while the primary path omits that projection. Build one final
  environment at the adapter bind seam so every launch surface uses the same
  projection.
- Untangle in the rewrite: `request.py` defaults `LaunchRuntime` to `DIRECT`,
  while ordinary production policy resolution rejects that surface and
  `tests/support/launch.py` recreates its preparation path. Use explicit
  production surfaces and exercise the real prepare/bind pipeline in tests.
- ops/spawn/api.py fork path: `inherited_skills` re-passes the source spawn's
  resolved skill names to Mars as extra skills, duplicating the profile's own
  loads (Mars dedupes, so it is harmless). Pass only the source's caller-requested
  extras once the snapshot records them separately.
- `system-prompt.md` is a launch input (Claude's `prompt_file_path`, OpenCode's
  `{file:}` agent prompt) but `write_projection_artifacts()` writes it, and each
  caller must run it after the spawn is persisted. Bind cannot own the write:
  `SpawnService.prepare()` binds before `lifecycle.start()` publishes the spawn dir,
  and `atomic_publish_dir` refuses an existing dir. To make it structural, have bind return the launch files and let
  `start_spawn` write them into the staged dir before publishing, or start the spawn
  before bind and roll the row back on composition failure.
- Launching-build shim (`cli_shim.py`) known limits: login shells rewrite PATH, so
  Codex (`bash -lc`) and a profile that resets PATH drop the shim from the model's shell.
  The shim pins the interpreter, not the build, so `uv tool upgrade`
  under a live session moves it. A dead shim is silent where a caller discards errors
  (Mars `context-autosync/run.sh`). `-P` breaks an uninstalled `cd src && python -m
  meridian` launch. Old `bin/<hash>/` directories are never cleaned.
- Launching-build shim, two more limits. The primary preview bind
  (`launch/__init__.py`) passes no `plan_overrides`, so it ignores `config.env`: a project
  `[env] PATH` that puts an old build first makes the real launch skewed where the preview
  saw none, and the refusal warning is silent. Passing the same config env to the preview
  closes it.
- Untangle in the rewrite: src/meridian/lib/launch/reference.py:135-160 increments the entry count once while filtering a blocked directory and again while rendering it. In a temporary directory containing only node_modules, max_entries=1 returned a truncation warning without printing the blocked entry; an ordinary single directory rendered normally. Count each displayed entry once in the rendering pass.
- Untangle in the rewrite: src/meridian/lib/launch/compiler.py:175-324 contains a local launch-parameter compiler with no in-repository callers; only its types and match_model_policy are imported by other modules. This leaves a second apparent policy-precedence implementation beside Mars launch-bundle. Do not port the unused compiler without a demonstrated consumer; retain the shared data and matching behavior only where needed, and check external imports before deleting the Python entry point.
- Untangle in the rewrite: src/meridian/lib/launch/command.py:48-57 carries a native reference-file injection path guarded by a capability that all five registered harnesses set false (src/meridian/lib/harness/adapter.py:102 and adapter declarations). This is an unexercised extension seam with no current native target; project references through the existing path until a supported harness needs native injection, then add the seam with that adapter. Verify external adapters before removing the Python field.
