matlab-dark-port

Port legacy MATLAB engineering code to Python through the legal deny-list gate — extract example data, audit branded names, build regression tests, and commit to nested repo correctly.

SKILL.md

GitHub

MATLAB Dark Port Workflow

Overview

Port legacy MATLAB engineering code (copyrighted, branded) into clean Python implementations that pass the legal deny-list scan. Extracts only methodology and numerical constants from the published standard — never copies function structure or variable names.

When to Use

  • User says "darkly port" or "dark port" MATLAB code
  • Legacy MATLAB with copyright headers, branded tool names, named authors
  • .legal-deny-list.yaml has entries blocking the original tool/firm/author

Procedure

Step 0 — Legal Audit BEFORE Writing Any Code

  1. Read .legal-deny-list.yaml to know BLOCKED patterns.
  2. Scan ALL MATLAB files with regex to catalog:
    # Function names
    re.findall(r'function\s+.*?=?\s*(\w+)\s*\(', content)
    # Branded strings (version headers, tool names)
    re.findall(r"VersionDetail.*?'(.*?)'", content)
    # Author names
    re.findall(r'(?:Author|By|Modified|Name)\s*[:=]\s*([A-Z][a-z]+ [A-Z][a-z]+)', content)
    
  3. Build a MATLAB→Python naming map. Present to user for approval.
  4. Any name containing a deny-list hit MUST be renamed.

Step 1 — Check Existing Python Implementation

The Python modules may ALREADY exist from a prior clean-room pass:

find */src/*/subsea/pipeline/free_span/ -name "*.py" | xargs wc -l

Check for NotImplementedError / pass — if real code exists, you only need to add the missing module + tests, not rewrite everything.

Step 2 — Extract Example Data from MATLAB

  • Read the "Setup" or "default generation" function — it contains hardcoded example inputs (pipe dimensions, current profiles, S-N curves, safety factors). This is the PRIMARY source of test vectors — it writes a sample .dat file with all default parameters and is often the only "worked example" available.
  • Read the "GetInputs" function for safety factor tables and parameter derivation.
  • These numerical values become regression test vectors.
  • NEVER copy variable names or function structure.
  • Expect no result files. Legacy MATLAB typically has no saved outputs — only .xlsm post-processing workbooks. Build tests from inputs + standard equations, not captured outputs.

Step 3 — Port Missing Modules Only

  • Write new modules with completely independent names.
  • Reference only the public standard (DNV-RP-F105 section numbers in docstrings).
  • Use standard Python patterns (dataclasses, scipy, numpy).

Step 3b — Resolve Cross-Module Import Dependencies

If the ported code depends on a module only available in the nested repo (e.g., digitalmodel.structural.fatigue.sn_curves), build a self-contained fallback module with _ prefix (private) inside the same package:

# _bilinear_sn.py — self-contained fallback
_DNV_SN_TABLE = {"F": {"air": {"A1": 1.73e11, "m1": 3.0, ...}, ...}}

def get_sn_curve(curve_class, environment="air"):
    """Built-in DNV-RP-C203 Table 2-1 lookup."""
    ...

Then in the consumer module, use try/except fallback:

def _load_sn_curve(self):
    try:
        from digitalmodel.structural.fatigue.sn_curves import get_dnv_curve
        return get_dnv_curve(self._inp.sn_curve_class)
    except (ImportError, ModuleNotFoundError):
        pass
    from ._bilinear_sn import get_sn_curve
    return get_sn_curve(curve_class=..., environment=...)

This pattern:

  • Works in both workspace-hub overlay AND digitalmodel venv
  • Uses the richer upstream module when available
  • Keeps the package self-contained and testable
  • Eliminates ALL skip markers from tests

Step 4 — Build Regression Tests

  • Create a pytest fixture using the extracted numerical values.
  • Guard tests that depend on uninstalled submodules:
    try:
        from digitalmodel.structural.fatigue.sn_curves import get_dnv_curve
        _HAS_STRUCTURAL = True
    except ImportError:
        _HAS_STRUCTURAL = False
    
    @pytest.mark.skipif(not _HAS_STRUCTURAL, reason="...")
    def test_fatigue_life_positive(self): ...
    
  • Verify safety factor tables match BOTH the standard AND the MATLAB defaults.
  • Try importing the module first — legacy code may have syntax errors.

Step 5 — Legal Scan + Commit

  1. bash scripts/legal/legal-sanity-scan.sh --diff-only → must PASS
  2. Double-check: grep -inE 'PATTERN1|PATTERN2' new_files → must be empty
  3. Commit tests in workspace-hub, implementation in digitalmodel (nested repo).

Pitfalls

1. Dual src/ paths (overlay vs nested repo)

workspace-hub has BOTH src/digitalmodel/... (editable install overlay, gitignored) AND digitalmodel/src/digitalmodel/... (real nested git repo).

  • New modules must be copied to BOTH paths for tests to find them.
  • git add in workspace-hub won't pick up src/digitalmodel/ (gitignored).
  • Implementation commits go to cd digitalmodel && git add && git commit.
  • Test commits go to workspace-hub root.

2. Zero-exceedance probability in Weibull fitting

MATLAB's round(1e8*(1-cumsum(occ)))/1e8 produces 0.0 for the last entry when occurrence probabilities sum to 1.0. Must filter P <= 0 before calling log() — otherwise math domain error.

3. Cross-module imports not available

SpanFatigueDamage depends on digitalmodel.structural.fatigue.sn_curves which may not be available in the workspace-hub overlay. Preferred fix: build a self-contained fallback module (Step 3b above) so tests never need skip markers. The _HAS_STRUCTURAL skip pattern is a temporary workaround only — replace it with the fallback module in the same session if possible.

3b. Environment-specific S-N curve behavior

DNV-RP-C203 seawater_cp curves have no fatigue limit (CAFL=0) per Sec 2.4.4. Tests asserting "zero damage below CAFL" must use IN_AIR environment, not SEAWATER_CP. The old single-slope code incorrectly applied the in-air CAFL to all environments.

4. Pre-existing syntax errors

Always try importing the target module BEFORE writing tests. Prior commits may have introduced syntax errors (e.g., double assignment on one line). Fix these first and include in the commit.

5. MATLAB safety factor naming swap

The MATLAB code names onset safety factors OPPOSITE to DNV convention:

  • MATLAB: onsetSafetyFactor_CF = 1.1, onsetSafetyFactor_IL = 1.3
  • Python (DNV convention): gamma_on_IL = 1.1, gamma_on_CF = 1.3 Document this mapping explicitly in test docstrings.

6. S-N curve transfer stress calculation

For bilinear S-N curves, the transfer stress is: S_transfer = (A2 / N_transfer)^(1/k2) Don't assume the result falls in a "typical" range — compute it from the actual parameters and assert around the computed value.

7. Git push for nested repos

git push in workspace-hub won't push the nested digitalmodel/ repo. Must explicitly: cd digitalmodel && git push origin main.

8. numpy 2.x compatibility

np.trapz was removed in numpy 2.x → use np.trapezoid. Compat shim:

_trapz = getattr(np, "trapezoid", None) or np.trapz

workspace-hub venv has numpy 2.x. Apply in both source and tests.

9. Post-implementation issue hygiene

After completing the port, the user expects:

  • Status comment on the parent issue with what was done + what's blocked
  • New child issue for any dependency gap (e.g., missing bilinear S-N support)
  • Bidirectional links between parent, child, and related issues
  • Follow-on issues for enhancement features not in scope (each with clear AC)
  • Close parent when all acceptance criteria met, even if follow-ons remain Use gh issue create + gh issue comment in the same flow.

10. VIV piecewise response: max current ≠ max damage

VIV amplitude models (F105 Fig 4-2, 4-5) are piecewise — the response peaks at an intermediate reduced velocity (Ur ≈ 7–8 for CF) and drops to zero past Vr_end (≈16). So the highest current can give LESS damage than a mid-range current if the reduced velocity overshoots the peak. Never assume "max current = worst case" in tests — instead find the actual worst bin:

max_single = max(assess(speed).damage_per_year for speed, _ in bins)
assert weighted_damage < max_single  # correct
# NOT: assert weighted_damage < assess(max_speed).damage_per_year  # WRONG

11. Multi-current probability-weighted damage pattern

The MATLAB approach loops over current bins and sums D_i × prob_i. Implement as a facade method that replaces current_velocity_ms via dataclass replace():

def assess_multi_current(self, current_bins):
    weighted_damage = sum(
        self._assess_at(speed).damage_per_year * prob
        for speed, prob in current_bins
    )

Return worst-case screening details (frequencies, amplitudes) from the bin that produced the highest unweighted damage, but use weighted damage for life.

12. JONSWAP spectral velocity integration

When computing wave-induced velocity from sea-state (Hs, Tp, depth):

  • JONSWAP gamma auto-selection: phi = Tp/sqrt(Hs), gamma = 5/exp/1
  • Dispersion: Newton-Raphson with deep-water initial guess k₀ = ω²/g
  • Velocity transfer: G(ω) = ω·cosh(k·z)/sinh(k·h), clip kh < 50
  • Spectral moments: M0, M2 via trapezoidal integration
  • Uw = 2√M0, Tu = 2π√(M0/M2)
  • Hs recovery from JONSWAP integration is ~15% off for high gamma — this is a known C_gamma normalisation artifact, not a bug.

Verification

# Tests pass
uv run python -m pytest tests/subsea/pipeline/test_*.py -v

# Legal scan clean
bash scripts/legal/legal-sanity-scan.sh --diff-only

# No deny-list patterns in new files
grep -inE '2H.?Offshore|2HSPANVIV|TwoH|AUTHOR_NAME' new_files