…slice 5)

The `tool_runner/dispatch.rs` match table previously narrowed every arm back to `Result<String, String>` via `.map_err(|e| e.to_string())` at the boundary, then stuffed the string into a hard-error `ToolResult` — discarding the `ToolError` kind that every submodule had already produced.

The table now yields `Result<String, ToolError>` natively and converts once, at the end, through `tool_result_from_typed`. That routes each error's `execution_status()` onto the `ToolResult` for every tool, generalising the #5984 fix (previously only the ACL-gated memory/wiki tools): a tool-level `PermissionDenied` (e.g. agent_send depth-exceeded) now surfaces as the soft `Denied` status, reported to the model but not counted toward the consecutive-hard-failure abort, instead of the stringly path's hard default.

Inline stringly arms map to their semantic variants: missing url/query -> `MissingParameter`, unknown tool -> `NotFound`, browser-subsystem-missing -> `Unavailable`. The handful of still-stringly downstreams (`browser_tools`, `web_fetch_to_file`, `channel::tool_channel_send`, MCP / skill providers) keep a `.map_err(ToolError::upstream_msg)` bridge pending their own slices.

Public `execute_tool` / `execute_tool_raw` signatures are unchanged (still return the `ToolResult` struct), so there is no downstream API change.

- Removes 67 boundary `.map_err(|e| e.to_string())` narrows
- Drops dispatch.rs from 10 `Result<String, String>` sites to 0
- Adds converter-boundary tests pinning the soft/hard status mapping and the typed wire strings
- Updates the now-stale "Unknown tool" assertion in tests/shell.rs to the typed `NotFound` phrasing
- Refreshes the error-contracts RFC migration order (slice 5 done) and the error.rs module doc

…mment

Comments referencing the current task slice number and superseded issue
IDs will rot as the codebase evolves. Trim them: drop the module-level
block comment in the test mod (test names are self-describing), shorten
the test helper comment and the `typed_not_found` inline note, and remove
the `shell.rs` change comment (assertions are already clear).

The invariant comment before `tool_result_from_typed` is reworded to
state the non-obvious WHY (soft vs hard abort counting) without the
issue numbers.