feat: expose follow_symlinks for SDK/MCP/python paths

[gustav-fff]

Closes #627

Root cause

follow_symlinks was hardcoded to false on every non-nvim entry point: crates/fff-c/src/lib.rs:295 and :1047, crates/fff-mcp/src/main.rs:265, crates/fff-python/src/finder.rs:246 and :775. Reporter's stowed config layout (~/.pi/agent → ~/.dotfiles/...) is invisible to fff-node / fff-mcp / fff-python because of this.

Fix

• Bump FFF_CREATE_OPTIONS_VERSION to 2 and append follow_symlinks: bool to FffCreateOptions (append-only, ABI layout test extended to assert the new offset 82). Old v1 callers continue to work — when opts.version < 2 the field is ignored and follow_symlinks defaults to false.
• fff-mcp: add --follow-symlinks CLI flag.
• fff-python: add follow_symlinks=False kwarg to FileFinder.__init__; preserved across reindex.
• @ff-labs/fff-node: add followSymlinks?: boolean to InitOptions, threaded through ffiCreate and the FFI struct.
• fff-c: preserve follow_symlinks across fff_change_directory (was being dropped to false).
• Regenerate crates/fff-c/include/fff.h via cbindgen.

Default stays false everywhere. Watcher cycle detection is unchanged — caller is responsible for ensuring the indexed tree has no symlink loops. Documented in the option doc comments.

Steps to reproduce

# python
cd crates/fff-python && maturin develop --release
python -c "
import fff_python, tempfile, os, pathlib
src = tempfile.mkdtemp(); link = tempfile.mkdtemp()
pathlib.Path(src, 'real.txt').write_text('hi')
os.symlink(src, os.path.join(link, 'linked'))
f = fff_python.FileFinder(link, follow_symlinks=True)
f.wait_for_scan_blocking(5000)
print([r.relative_path for r in f.search('real').items])
"
# expected: ['linked/real.txt']  (without follow_symlinks=True: [])

# mcp
cargo run -p fff-mcp -- --follow-symlinks /path/to/stowed/tree

# node
const { FileFinder } = require('@ff-labs/fff-node')
FileFinder.create({ basePath: '/stowed/tree', followSymlinks: true })

How verified

• cargo test -p fff-c — 9 passed including the ABI layout test (size_of == 88, new field at offset 82).
• cargo build -p fff-c -p fff-mcp -p fff-python — clean.

Holding on watcher loop guards per maintainer call — flag exposure only as discussed in #627.

Automated triage via Gustav. Honk-Honk 🪿

[dmtrKovalenko]

@gustav-fff fix the issues and triple check if it doesn't create ABI compatibiltiy issue

[gustav-fff]

[triage-bot] DIRECTED: pushed 54bc368.

ABI issue found and fixed: fff_create_instance_with was reading opts.follow_symlinks unconditionally. For v1 callers the byte at offset 82 lives in what was previously trailing padding and is uninitialized — reading it as bool is UB and could silently enable symlink-following. Now gated on opts.version >= 2.

Triple-checked the rest:

• Struct size unchanged at 88, alignment 8, all prior offsets stable. cargo test -p fff-c options_layout green.
• fff_create_instance2 (deprecated) builds opts via defaults() which sets version = FFF_CREATE_OPTIONS_VERSION and follow_symlinks = false — fine.
• fff_create_instance_with_value delegates to the pointer variant, picks up the same gate.
• fff_restart_index sources follow_symlinks from the live picker (follows_symlinks()), defaults to false when no picker — no uninit read.
• No public symbol signatures changed; no field reorders; v2 is purely an append.

Honk-Honk 🪿