Adds streaming RPC forwarding to collab so guests can call
`GetInitialGraphData` and `SearchCommits` against a remote host project.
Previously these requests had no forwarder registered on the server and
would fail when invoked by a guest.

This mirrors the existing single-response forwarding pattern with new
analogues:
- `StreamResponse<R>` + `MessageContext::forward_request_stream`
- `Server::add_request_stream_handler`
- `forward_read_only_project_stream_request`, registered for both
messages

Also hardens both the unary and stream handlers to send
`respond_with_error` when a handler returns `Ok` without sending/ending
a response, so the client doesn't hang waiting for a reply that will
never arrive.

I added git graph collab integration tests for this as well. 


Self-Review Checklist:

- [x] I've reviewed my own diff for quality, security, and reliability
- [x] Unsafe blocks (if any) have justifying comments
- [x] The content is consistent with the [UI/UX
checklist](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist)
- [x] Tests cover the new/changed behavior
- [x] Performance impact has been considered and is acceptable

Closes #55954

Release Notes:

- N/A

Pasting text that triggers slash-command completions caused a reentrant update of MessageEditor: the paste action ran inside MessageEditor::update, which entered Editor::update to do the paste, which synchronously fired trigger_completion_on_input, whose provider then called slash_autocomplete_invoked and tried to MessageEditor::update again.

Defer the SlashAutocompleteOpened emit via cx.defer so it runs after the surrounding update tree unwinds.

Code had been assuming (erroneously, but understandably) that
LlmApiToken::acquire would give them a valid token.

This is not true, as those tokens expire and you must call refresh explicitly.

Add some helpers to do the retry for you, and rename acquire to cached to be
clearer about the intent.

Co-authored-by: Danilo Leal <daniloleal09@gmail.com>

Builds on AI-218 (which discovered and loaded skill paths) by adding
frontmatter parsing. The `Skill` struct now carries `name`,
`description`, and `disable_model_invocation` in memory. Skill bodies
stay on disk \u2014 they're read on demand by `read_skill_body` only when
a skill is actually materialized for the model.

- Add `parse_skill_frontmatter`: parses YAML frontmatter and validates
  per the agent-skills spec (`name` matches `[a-z0-9-]{1,64}`,
  `description` is 1\u20131024 chars).
- Add `load_skill_frontmatter`: chunked reader. Uses `fs.open_sync` +
  inline 4KB chunks, stops as soon as the closing `---` is found (or
  `MAX_SKILL_FILE_SIZE` is hit). Reads synchronously inside an async
  function; production callers already wrap `load_skills_from_directory`
  in `cx.background_spawn`, so blocking happens on the background
  executor. See the inline comment for why `smol::unblock` doesn't fit
  (it panics under `TestAppContext`).
- Add `read_skill_body`: `fs.load` + extract body. Public API but
  unused in this PR \u2014 the user-facing PR consumes it.
- `load_skills_from_directory` now returns
  `Vec<Result<Skill, SkillLoadError>>` so a future PR can surface
  parse errors in the UI. `build_project_context` filter-maps Ok
  variants for now, logging errors at `log::warn` level.

Test fixtures in agent.rs updated from placeholder bytes to valid
frontmatter; assertions strengthened from path-equality to checking
parsed `name` / `description`.

Have the editor push the cursor's buffer path to Project::set_active_path

on local selection changes and on focus. This drives the existing

ActiveEntryChanged consumers (project panel auto-reveal, Copilot DidFocus,

edit prediction recent paths) for multibuffer items, where the workspace

resolver only sees a None project_path.

Singleton editors take the same path; set_active_path dedups against the

workspace's own pane-event-driven calls.

Release Notes:

- Fixed the project panel not auto-revealing the file containing the

  cursor when navigating between excerpts in a multibuffer (project diff,

  diagnostics, project search, etc.).

`zed --new ssh://host/path/to/file` only created a new window on the first call. Subsequent invocations silently reused the existing SSH workspace for the same host, which then tried to open the new path against the existing worktree and surfaced a `DevServerProjectPathDoesNotExist` popup when the path didn't belong to any open worktree.

The CLI correctly translated `--new` to `OpenBehavior::AlwaysNew`, but the URL branch of `handle_cli_connection` dropped the `open_behavior`. `handle_open_request` then called `open_remote_project` (and `open_paths_with_positions` for `file://`) with `OpenOptions::default()`, which is `WorkspaceMatching::MatchExact`, so any existing SSH window for the same host won the match. The issue applies to `file://` URLs and `-a` / `-e` / `--reuse` for URL-shaped arguments in general; `--new` was just the most visible symptom.

The fix is making sure the translation of `OpenRequest` is uniform across plain paths and URL-shaped arguments.

Closes #52679.

Release Notes:

- Fixed `zed --new ssh://host/path` reusing an existing SSH window instead of opening a new one

The data structure for a loaded skill is now just the paths \u2014 source,
directory path, and SKILL.md path. Parsing the YAML frontmatter
(name, description, disable-model-invocation) and the body comes in a
follow-up PR.

This roughly halves the diff size of the PR by dropping:
- `parse_skill`, `extract_frontmatter`, `validate_name`, `validate_description`
- `SkillMetadata`, `SkillSummary`, `SkillLoadError`
- `MAX_SKILL_FILE_SIZE`, `MAX_SKILL_DESCRIPTIONS_SIZE`
- `load_single_skill` (no per-file parsing path)
- `merge_skills` and its test (dedup is by parsed name; not in scope)
- `SkillLoadingError`, `SkillLoadingErrorsUpdated`, and the
  associated error emission in `maintain_project_context` (without
  parsing, there's nothing to error on)
- The `skill_loading_errors` field on `ProjectState`
- The `anyhow`, `serde`, `serde_yaml_ng`, and `serde_json` deps of
  `agent_skills`
- The workspace `serde_yaml_ng` entry (no other workspace member uses
  it on this branch)
- All parsing tests

`load_skills_from_directory` now returns `Vec<Skill>` directly,
`build_project_context` returns `Task<(ProjectContext, Vec<Skill>)>`,
and `state.skills` ends up as the concatenation of global skills +
each worktree's project-local skills (no dedup since we have no
parsed name to key on).