fsck: BAD_FILEMODE diagnostic broken / never triggers

While looking at git's handling of invalid objects, I stumbled upon git apparently not caring much about file mode values, except for the following requirements:

- must be non-empty (badTree)
- must be octal (badTree)
- should not be zero-padded (zeroPaddedFilemode)

Looking at tree-walk this is obvious from decode_tree_entry: it calls canon_mode, which

- checks the S_IFMT exactly, and falls back on 160000 (S_IFGITLINK)
  for cases other than S_ISREG, S_ISLNK, or S_ISDIR
- returns the corresponding S_IF (with the access modes unset) for all 
  cases other than S_ISREG, for which it returns 755 if S_IXUSR is set,
  otherwise 644

However looking at fsck_tree, it does have a fair amount of code to validate entry modes and a dedicated message id for this (BAD_FILEMODE / badFilemode), it even has a code path for legacy entries with S_IWGRP set (extensively documented under `git fsck --strict`).

I guess, over time mode canonicalisation has slowly creeped earlier the tree-parsing code, and (seemingly for several years) it has been occurring before "git fsck" gets tree entry information, so git fsck simply can not see invalid entry modes?

Reproduction:

- create an empty git repository
- create tree for (or with) any possible entry type value
- run `git fsck --strict --no-dangling`

Expected:

all the trees except the 5 with valid filemodes should be flagged

Observed:

Checking object directories: 100% (256/256), done. notice: HEAD points to an unborn branch (master) notice: No default references

Script to generate one tree for each filemode between 1 and 777777 (inclusive), can be run from a (non-bare) repository root, beware that the resulting repository takes about 1GB.

from hashlib import sha1 from pathlib import Path from zlib import compress

def hash_object(type, content):
    obj = b"%s %d\0%s" % (type.encode(), len(content), content)
    oid = sha1(obj).hexdigest()

    e_dir = GIT_ROOT / 'objects' / oid[:2]
    e_dir.mkdir(exist_ok=True, parents=True)

    e = e_dir / oid[2:]
    if not e.exists():
        e.write_bytes(compress(obj))

    return oid

# technically S_IFMT caps out at 200_000 - 1, but get_mode doesn't
# actually have a limit
MAX_ENTRY = 0o777_777
if __name__ == '__main__':
    GIT_ROOT = Path.cwd() / '.git'
    assert GIT_ROOT.is_dir()
    
    blob_hex = hash_object('blob', b"")
    blob_oid = bytes.fromhex(blob_hex)
    print("empty blob", blob_hex)

    tree_hex = hash_object("tree", b"100644 empty_blob\0" + blob_oid)
    tree_oid = bytes.fromhex(tree_hex)
    print("base tree", tree_hex)

    for mode in range(1, MAX_ENTRY+1):
        print(f"\r{mode:06o}/{MAX_ENTRY:o}", end=' ')
        tree = b"%o x\0%s" % (
            mode,
            tree_oid if (mode & 0o170_000) == 0o040_000 else blob_oid,
        )
        h = hash_object("tree", tree)
        print("=", h, end='', flush=True)