perf: scan the workspace once per create instead of once per port

[okazu-dm]

Summary

• sprout create walked the entire workspace once per {{ auto_port() }} placeholder. It now walks it once per create, and the walk itself skips dependency and cache directories
• On the workspace that motivated this (~490k files under .sprout/, 8 auto-assigned ports), the .env generation step went from 97s to 2s; a single walk went from 10.0s to 0.66s
• Port results are unchanged: on a real workspace with 661 .env files, the old and new walks return identical file sets and identical port sets

Problem

Two independent causes, both measured with per-phase timestamps:

1. find_available_port() called get_used_ports() on every call, so .sprout/ was walked once per placeholder — 9 walks for a template with 8 ports, at ~11s each.
2. Every walk descended into node_modules, .venv, .git and friends, which held 77% of the files under .sprout/.

Changes

• find_available_port(used_ports=None) accepts a pre-computed port set. parse_env_template() builds that set once per file and reuses it, so get_used_ports() runs once per create instead of once per port.
• iter_env_files() replaces Path.rglob("*.env") with os.walk(), pruning SCAN_EXCLUDED_DIRS in place.
• The scan is deferred until the first {{ auto_port() }} is reached, so a template without one never walks the worktrees — and never shells out to git rev-parse.

Correctness notes

Three ways this can go wrong, and how each is handled:

• A worktree can be named after an excluded directory. sprout create venv (or fix/venv) puts a real worktree behind an excluded name. Pruning by name alone made its .env invisible and handed its ports out again. is_port_available() is not a safety net here: it only sees a conflict while the other stack is running, so the collision surfaces later as a failure to bind. An excluded name is still walked when the directory holds a .git entry, which every worktree root does.
• used_ports replaces the scan rather than adding to it. An empty set means "nothing is taken"; pass None to have the function scan. Both docstrings spell this out, because a partial set silently gives up collision detection against the other worktrees.
• Non-regular files. Dropping the is_file() guard let a fifo named .env block read_text() indefinitely. The guard is back, applied only to names that already end in .env, so it costs one stat per candidate.

Test Plan

• make all-checks green: ruff, mypy, 103 tests (up from 70)
• Old vs new walk compared on a real 490k-file workspace: identical file sets (661 files) and identical port sets (48 ports)
• Parity checked for dotfiles, local.env-style names, directories named *.env, dangps, symlinked worktrees and permission errors
• Mutation tested: 11 mutations — prune removal, dropped worktree-root escape, unconditling in both functions, per-file rescan, shrunken exclude list, exact-name match, missingfile_ports.add(), dropped is_file() — each kills at least one test
• End-to-end sprout create on the original workspace: .env generation 97s -> 2s

[laysakura]

Thx. I'll merge it and upgrade to 0.9.0.