fix(discover): detect extensionless scripts by shebang

[ibaldr89]

Summary

• fall back to a bounded first-line shebang probe when filename/special-filename/extension detection returns unknown
• recognize direct interpreter paths plus /usr/bin/env and env -S for common supported scripting languages
• preserve filename/extension precedence and fail closed on malformed, NUL-containing, truncated, or unreadable first lines
• add production-discovery regressions for valid, conflicting, malformed, binary, env-assignment, and boundary cases

Fixes #1199

Why

Executable source scripts commonly omit a filename extension. Before this change, files such as:

#!/usr/bin/env python3

def extensionless_probe():
    return 42

were silently absent from the graph even though Python is supported. Full indexing did not change that because language detection was filename-only.

Behavior

Existing detection remains authoritative. Shebang detection runs only when cbm_language_for_filename() returns CBM_LANG_COUNT.

Supported interpreter basenames:

• Python: python, python2, python3, dotted numeric versions such as python3.12
• shell: sh, bash, dash, ksh, zsh
• JavaScript: node, nodejs
• ruby, perl, php, lua

The probe reads at most 255 bytes plus one EOF/truncation probe byte, rejects embedded NUL, read errors, overlong first lines, unsupported env option/assignment shapes, and arbitrary prefix matches such as python-wrapper.

Related work checked

• feat(discover): user-defined extension→language mappings (C port) #73 and its documentation follow-up docs(config): document extra_extensions values and skip behavior #487 add user-defined extension mappings, but those mappings require a dotted extension and cannot represent extensionless files.
• Option to index new filetypes #962 concerns path/glob coverage and reporting for skipped files, not interpreter-based language detection.
• Windows: Claude hook installer creates extensionless hook files causing OpenWith dialog in Cursor #929 / fix(install): Windows hook scripts get .cmd form; legacy files removed #1057 concern extensionless Windows installer hook scripts, not source discovery.

No existing issue or PR found in the checked open/closed history implements shebang-based language fallback for extensionless source files.

Testing

• New tests added
• Focused discovery suite: 107 passed under ASan/UBSan
• CI lint stack: clang-format 20 + cppcheck 2.20 + NOLINT policy passed
• GitHub Actions full matrix: 20/20 checks passed, including Linux x86/ARM, macOS, Windows, TSan, CodeQL, DCO, and lint
• Real-repository regression: 7/7 extensionless entrypoints indexed versus 0/7 before; expected function and call graph became queryable
• Independent exact-tree review: approved for commit

Full suite comparison in the same isolated Docker environment:

• candidate: 6198 passed, 0 failed
• clean main: 6180 passed, 0 failed
• both runs exit non-zero on the same pre-existing 61-byte LeakSanitizer finding in tests/test_httpd.c:358; no candidate-only failure was introduced