One implementation of string slice/substring/replace; check inline code passed through process.run

[b-macker]

Summary

Round 3 of dogfooding (the repo-sentinel project) turned up two problems. This PR fixes both.

1. String methods disagreed between engines, and one hung

Round 3 reported F-007, the string.slice hint, as "not fixed". It had run a build without #261, which does suggest substring. Checking it turned up something bigger:

• s.slice(...) worked as a method, while string.slice(...) didn't exist as a function.
• The method forms of three string operations gave different results in the VM and the tree-walker.

Each operation had four hand-written copies: VM method dispatch, two tree-walker method paths, and the string module.

Call	VM	Tree-walker
s.replace("-", "+")	a+b-c (first match only)	a+b+c
s.substring(3, 1)	""	"lo" (a length calculation underflows)
s.slice(-3)	llo	hello (treated as substring)
"ab".replace("", "+")	+ab	hangs

The last row is the serious one. The tree-walker's replace didn't guard against an empty pattern. An empty string matches at every position, so it kept inserting forever while memory grew. --timeout can't stop it, because the loop never returns to the interpreter, and the REST API runs the tree-walker.

2. Inline code in process.run skipped the code checks

repo-sentinel reported every C++ file as 0 bytes. The cause was a Python helper that swallowed every error:

except Exception:
    return 0

The project's own code_quality.no_incomplete_logic (HARD) blocks that exact code in a <<python>> block and in codegen.run(). repo-sentinel ran it as process.run("python3", ["-c", code]) instead, and it went through with no check at all. NAAb already had the rule that would have caught the bug.

Changes

• New include/naab/string_ops.h with substring, slice and replaceAll, called from all four places:
  • The string module functions already agreed with each other in every case, so their behaviour is the reference.
  • replace replaces every occurrence; an empty pattern leaves the string unchanged.
  • substring clamps out-of-range indices and returns "" for an empty or reversed range.
  • slice follows JavaScript: negative indices count from the end.
• string.slice(s, start[, end]) is now a module function, behaving the same as the method.
• string.substr keeps its hint pointing to substring, because JavaScript's substr takes a length rather than an end index.
• process.run inline-code check (src/stdlib/process_impl.cpp): when the command is an interpreter given inline code, that code goes through checkPolyglotBlock(), the same check <<python>> blocks and codegen.run() use. Recognised:
  • Interpreters and flags:
    • python/pypy -c
    • node -e/-p/--eval/--print, also --eval=
    • ruby -e
    • perl -e/-E
    • php -r
    • sh/bash/dash/zsh/ksh -c
  • Also matched through a directory prefix, .exe, a version suffix (python3.12), and combined short flags when the code flag comes last (-Ic, -ec).
  • Side effect: languages.allowed/blocked now apply to that inline code too.
  • Not covered: a script file (python3 s.py) isn't scanned, and only the outer interpreter is recognised.
• CLAUDE.md: notes for both changes.

Test Plan

• String fixes:
  • New differential probe tests/differential/corpus/string_methods.naab. The corpus only exercised the module form of these functions, which is why none of this showed up. Differential run: 57 passed, 0 failed.
  • tests/parser/test_dogfood_hints.sh: 21/21 pass. S-06 checks string.slice works; E-01 checks empty-pattern replace under a 10 s limit.
  • Against the previous build, S-06 and E-01 fail in both engines; E-01 on the tree-walker fails by being killed at 10 s.
• process.run check:
  • New tests/security/test_process_run_inline_gate.sh: 14/14 across both engines.
  • Every blocking case checks that the code's marker file was never written, not just the exit code.
  • Against the previous build it fails only the two blocking cases (PI-01, PI-03), in both engines.
  • Controls:
    • PI-00: the <<python>> block is blocked under this config, so the config does exercise the check.
    • PI-02, PI-04, PI-05: harmless python3 -c, harmless sh -c and echo -c <text> still run.
    • PI-06: a script file is out of scope and still runs.
• Existing callers: the python3 -c and sh -c calls in living-script, living-script_v2, living-script_extended, hivemind and hivemind_governed run unchanged under their own govern.json, on both builds.
• Leak check: 874 passed, 0 failed. The new code adds no error text of its own; it passes on checkPolyglotBlock's messages.
• Full suite: a fresh run including the new test is pending. I'll post the result here.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ELUfjXZvx8kzXo1UJjrAhC

[github-actions]

NAAb Governance Report

Metric	Count
Files checked	16
Passed	16
Failed	0

All governance checks passed!

_{Generated by NAAb Governance Engine v4.0}

[b-macker]

Local results at 325b765 (includes the process.run inline-code gate):

• Full suite: 447 total, 381 passed, 52 error-behavior, 12 needs-tree-walk, 2 missing-executor, 2 unexpected.
  • The 2 unexpected are test_require_governance_gov007.sh and test_platform_fixes.sh.
  • Both are caused by a stray /govern.json on this machine, and the pre-change baseline (ce0ca24) had the identical result.
  • The newly registered test_process_run_inline_gate.sh passed.
• Error-message leak check: 874 passed, 0 failed.
• Differential v2: 57 passed, 0 failed.

CI on 325b765 is green, including build-windows.

---

Generated by Claude Code