fix(toolpath-convo): strip leading slash in unified_diff headers

[eliothedeman]

Closes #36

Summary

unified_diff(path, ...) previously emitted --- a//abs/file.rs / +++ b//abs/file.rs whenever path already started with / (common for Claude's tool_use.input.file_path, which carries absolute paths like /Users/.../src/login.rs). Git-style a/ and b/ prefixes already denote the repo root, so the doubled slash breaks patch(1) and any other tool that parses the header.

The fix trims a single leading / before splicing into the header. The toolpath-claude test that asserted on the broken form ("--- a//src/login.rs") has been updated to match — it had been canonicalising the bug.

Changes

• crates/toolpath-convo/src/derive.rs::unified_diff — strip leading / via trim_start_matches('/'); doc comment updated to explain the rationale.
• crates/toolpath-convo/src/derive.rs — add test_unified_diff_strips_leading_slash_on_absolute_path and test_unified_diff_preserves_relative_path regression tests.
• crates/toolpath-claude/src/derive.rs::test_derive_path_edit_tool_emits_unified_diff — assert the correct headers and explicitly reject any a// substring.

Risk

Low. The frontend diff renderer (toolpath-desktop) splits on lines and does not parse the header, so no UI breakage. External consumers piping .path.json diffs into patch(1) benefit directly. No public API change; no version bump.

Test plan

• cargo test -p toolpath-convo -p toolpath-claude
• cargo clippy -p toolpath-convo -p toolpath-claude -- -D warnings

[github-actions]

🔍 Preview deployed: https://fb9414f9.toolpath.pages.dev

[eliothedeman]

Summary

Targeted, correct fix for #36. unified_diff now strips the leading / so absolute paths stop emitting --- a//abs/path.rs. Both --- and +++ headers use the trimmed display, and the misleading Claude test is corrected rather than deleted.

Notes

• crates/toolpath-convo/src/derive.rs:381 — trim_start_matches('/') strips all leading slashes, not just one. The issue only spec'd stripping a single /, but this is the safer choice (nonstandard //foo inputs also normalize) and test_unified_diff_strips_leading_slash_on_absolute_path enforces the !contains("a//") invariant. Worth noting in case anyone expects Path::new-style "preserve // root" semantics, but I'd keep it as-is.
• Edge cases behave reasonably: empty string and bare / both produce --- a/\n+++ b/\n — same as before for the empty case, strictly better for /. Relative paths pass through untouched (covered by test_unified_diff_preserves_relative_path).
• crates/toolpath-claude/src/derive.rs:1619 — good that the updated test asserts both headers plus the negative !raw.contains("a//"); that's what keeps the canonicalised-bug from reappearing.
• Scope is clean: no version bump (correct — public signature and semantics for the common relative-path case are unchanged), no CHANGELOG entry (consistent with the file's version-scoped pattern), no refactoring drift.
• Risk: a downstream consumer that specifically parses a//abs/... would break, but that already means they worked around our bug. The PR body's patch(1) framing is right — this makes output correct per git conventions.

Verdict

LGTM (commenting rather than approving — GitHub blocks self-approval). Matches the issue spec, has regression coverage on both sides of the crate boundary, and the doc comment on unified_diff now explains why the trim happens so a future reader won't undo it.