feat(pwd): add rtk pwd for execution-context validation (#2148)

[oluies]

Summary

Adds rtk pwd — a cheap execution-context validation command, motivated by #2148 (see my investigation comment).

Agent harnesses (Claude Code, Cursor, …) track their shell cwd out-of-band and can silently drift to the wrong checkout when git worktrees are involved — commands the agent believes run in a linked worktree actually run in the main checkout, and rtk faithfully reports the "wrong" checkout's state (or commits there). rtk pwd gives agents and hooks a one-shot way to assert where commands will actually execute before writing:

$ rtk pwd                       # in a linked worktree
/repo/.worktrees/feature
branch: feature
worktree of: /repo (branch develop)

$ rtk pwd                       # in a main checkout, subdirectory
/repo/src/cmds
branch: develop
repo root: /repo

$ rtk pwd                       # outside git
/some/dir
(not in a git repository)

It also surfaces logical $PWD vs physical getcwd() mismatches — the innocent explanation behind "pwd and rtk proxy pwd disagree" reports (shell builtins print the symlink-preserving $PWD; child processes only ever see the physical cwd):

note: shell $PWD is a symlinked path to the same directory: /tmp/link
warn: shell $PWD differs from physical cwd: /other/path

Design

• Pure in-process FS walk — no child process, no dependency on a git binary; parses .git dir/file, the gitdir: pointer, commondir (with a path-based fallback when absent), and HEAD (branch or detached short hash) directly. Keeps startup well within the <10ms budget.
• Never fails the user — malformed metadata degrades to (unknown) / (unresolved gitdir pointer); only an unreadable cwd is an error.
• Token-light — 2–4 short lines; tracked via TimedExecution like other info commands.
• Not added to is_operational_command (diagnostic command, invoked explicitly — fail-open per the whitelist comment), and no hook rewrite rule (pwd builtin output must stay untouched).

Tests

10 hermetic unit tests with hand-crafted .git fixtures (no git binary needed): main checkout, subdirectory, linked worktree, missing commondir fallback, detached HEAD, malformed gitdir pointer, symlink note / differing-$PWD warn / equal-paths cases.

cargo fmt --all ✓ · cargo clippy --all-targets ✓ (0 warnings) · cargo test --all ✓ 2006 passed, 0 failed.

Verified end-to-end against the real binary in a linked worktree, main checkout, subdirectory, non-repo dir, and through a symlinked cd.

[oluies]

Salut @aeppling ! Petit bonjour de Suède 🇸🇪

Tu peux jeter un œil à cette PR (et à l'analyse dans #2148) quand tu as un moment ?

Question de fond : est-ce que cette dérive de cwd avec les worktrees est un problème déjà résolu ailleurs (un autre outil, une autre approche qu'on devrait reprendre), ou est-ce que c'est simplement — comme on dit en Suède, « skit bakom ratten » — de la merde derrière le volant, c'est-à-dire le conducteur (l'agent) qui se trompe de répertoire, et pas rtk ?

Mon enquête penche pour la deuxième option : rtk ne touche jamais au cwd (audit du code + 14/14 checks sur un repro déterministe), d'où ce rtk pwd comme garde-fou léger pour que l'agent puisse valider son contexte avant d'écrire.