directory entries - files added or deleted locally are invisible to `status`/`diff`/`pin`; `pin` says "nothing to pin" and a fresh install silently loses the added files

[eljulians]

Component: crates/cli (status/pin/diff for dir entries) | Severity: Medium-High (leads to data loss via false confidence) | Status: Reproduced

All dir-entry comparisons iterate the cache file list. A file the user deleted locally reads as "not modified" (missing installed file → return Ok(false)), and a file the user added is never examined at all. So: status clean, diff "clean - no local modifications", pin "matches upstream - nothing to pin" (writes no patch). The user reasonably believes their customization is captured; a fresh clone+install won't have the added file, and install --update resurrects the deleted one.

Where: crates/cli/src/commands/status.rs:29-32; crates/cli/src/commands/multi_target.rs:54-59; crates/cli/src/commands/pin.rs:101-118 (also deletes patches for files missing from the representative).

Repro:

# installed dir skill 'd' with files SKILL.md + helper.md
rm .claude/skills/d/helper.md && echo notes > .claude/skills/d/notes.md
skillfile status   # clean
skillfile diff d   # 'd' is clean - no local modifications
skillfile pin d    # 'd' matches upstream - nothing to pin

Expected: additions/deletions reported as modifications; pin either captures them (add-file patches / deletion records) or errors "cannot pin added/deleted files" so the user knows.
Actual: fully invisible.

Fix sketch: in the dir comparison, walk the union of cache and installed file sets; report installed-only files as added and cache-only-with-missing-installed as deleted; make pin at minimum refuse with a message listing them (full capture support can come later).

Tests:

1. Unit (crates/cli/src/commands/multi_target.rs or status.rs): cache={a}, installed={a,b} → modified; cache={a,b}, installed={a} → modified.
2. Functional (tests/cli.rs): the repro; assert status marks the entry and pin does not claim success.