fix(sc8): count only the walk's own CPU time against its runtime budget - #828

fix(sc8): count only the walk's own CPU time against its runtime budget

(cherry picked from commit 182a475)
Signed-off-by: yashrajbasav <[email protected]>

Since NVIDIA#828, the shipped-bytecode walk counts only its own CPU time against
the 5-second budget, and the workflow wall-clock deadline bounds elapsed
time. Word the resource-bounds row the same way as the YARA rule-load row.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Narendran Raghavan <[email protected]>

The install docs, the CI recipe and the missing-scanner hint pinned
SkillSpector v2.12.0. That release can make the Tier 1 security check
INCOMPLETE, or fail it, for skills with no real issue:

- valid Python and Perl helpers (for example a `len(x) < 12` comparison
  followed by more code, or a Perl `eval { require $file; 1 }` block) are
  recorded as `static_parse_limit` or `obfuscated_instruction_text`
  ledger exceptions, with a High AE1 at each line that names the helper
  (NVIDIA/SkillSpector#777);
- the SC8 bytecode walk charges time spent waiting behind other analyzers
  to its budget and can stop with a static `runtime_limit`
  (NVIDIA/SkillSpector#828);
- visible text between two HTML comments raises a High P2
  hidden-instruction finding (NVIDIA/SkillSpector#452).

SkillSpector v2.13.1 contains all three fixes. Move every install spec to
v2.13.1, keep the rule that docs and code name exactly one release, and
keep the Claude 5.5 caveat: v2.13.1 still lacks the newer structured-output
support, so SKILLSPECTOR_MODEL must still name a supported model when
SKILL_EVAL_LLM_MODEL is a Claude 5.5 model. The SkillSpector model
defaults are unchanged.

Add a troubleshooting entry for false incomplete helper scans on older
releases, and a captured v2.13.1 --no-llm report (from
tests/fixtures/skills/simple) to the captured-report acceptance test so the
strict completeness validator is exercised against the pinned release.

Signed-off-by: Narendran Raghavan <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>

* chore(deps): pin SkillSpector to v2.13.1

The install docs, the CI recipe and the missing-scanner hint pinned
SkillSpector v2.12.0. That release can make the Tier 1 security check
INCOMPLETE, or fail it, for skills with no real issue:

- valid Python and Perl helpers (for example a `len(x) < 12` comparison
  followed by more code, or a Perl `eval { require $file; 1 }` block) are
  recorded as `static_parse_limit` or `obfuscated_instruction_text`
  ledger exceptions, with a High AE1 at each line that names the helper
  (NVIDIA/SkillSpector#777);
- the SC8 bytecode walk charges time spent waiting behind other analyzers
  to its budget and can stop with a static `runtime_limit`
  (NVIDIA/SkillSpector#828);
- visible text between two HTML comments raises a High P2
  hidden-instruction finding (NVIDIA/SkillSpector#452).

SkillSpector v2.13.1 contains all three fixes. Move every install spec to
v2.13.1, keep the rule that docs and code name exactly one release, and
keep the Claude 5.5 caveat: v2.13.1 still lacks the newer structured-output
support, so SKILLSPECTOR_MODEL must still name a supported model when
SKILL_EVAL_LLM_MODEL is a Claude 5.5 model. The SkillSpector model
defaults are unchanged.

Add a troubleshooting entry for false incomplete helper scans on older
releases, and a captured v2.13.1 --no-llm report (from
tests/fixtures/skills/simple) to the captured-report acceptance test so the
strict completeness validator is exercised against the pinned release.

Signed-off-by: Narendran Raghavan <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>

* chore(ci): install SkillSpector v2.13.1 in the Tier 1 image

The install docs and the missing-scanner hint name SkillSpector v2.13.1,
but the Tier 1 CI image still installed the v2.12.0 source archive
(c7958a3), so CI scans kept the false incomplete and false High findings
that v2.13.1 fixes.

Point ci/requirements/skillspector.in and the hash-locked
skillspector.txt at the archive of the v2.13.1 tag commit
(f3b25697d76256998ca9d42c198d256258d4c035) and update the
THIRD_PARTY_NOTICES row. v2.13.1 declares the same dependencies as
v2.12.0, so only the archive URL and its hash change; the documented
uv pip compile command gives the same lock.

Extend test_skillspector_install_instructions_pin_one_release so the CI
lock's archive commit and the notices row must match the documented
release, and mention the image in the CHANGELOG entry.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Narendran Raghavan <[email protected]>

* chore(deps): pin SkillSpector to v2.13.3

Move the install instructions, Tier 1 CI lock (commit 38a1dfe, archive sha256
recomputed), third-party notice, and captured --no-llm fixture from v2.13.1
to v2.13.3.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Signed-off-by: Narendran Raghavan <[email protected]>

---------

Signed-off-by: Narendran Raghavan <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>