I built a crypto-agility scanner, then found a detection gap in it worth sharing

Most crypto-inventory tools assume one clean, modern codebase. African fintech infrastructure rarely looks like that — a COBOL or legacy Java core, a modern API middleware layer bolted on top, a PHP admin portal, a mobile client, often four different teams with four different crypto postures. CryptoRadar scans all of it together.
The approach: deliberately regex-based rather than full AST parsing. A scanner that works today across Java, Kotlin, COBOL, JS/TS, Python, PHP, Go, C#, and config files beats a perfect-but-single-language one for a first-pass inventory — especially when a COBOL core is in scope and no mainstream AST tooling covers it well.
The bug worth sharing: Java names signature algorithms as one concatenated token — SHA256withRSA, SHA1withECDSA — with no separator between the hash and the algorithm. Standard word-boundary regex matching cannot match inside that token from either side. I verified this empirically against six real compound-algorithm strings before trusting the fix: all six produced zero findings under the old pattern set. This is arguably the single most common way Java names a signing algorithm, so it was a real coverage gap, not an edge case. Fixed with a dedicated compound-token signature rather than patching the existing patterns with lookbehind hacks.
The other fix: comment-masking. A // comment describing legacy crypto fired the same finding as the real code beneath it — in one case, a file with no actual legacy crypto at all produced a pure false positive sourced entirely from a comment. Built a single-pass, offset-preserving masker that strips comment content but leaves string literals untouched, so a hardcoded secret inside a quoted string is still caught.
Real results: ran the updated scanner against a production-shaped fintech codebase (11 microservices, 676 files). Before: grade F, 26.4%, 10 Critical findings (a hardcoded secret shared across ten services). After rotating that secret and re-scanning: grade D, 31.4%, zero Critical findings — confirmed as genuine remediation, not a scanner artifact, by independently checking the updated scanner still detects the original secret pattern when present.
Try it:
pip install "git+https://github.com/bennybriel/cryptoradar.git@v0.2.0"
cryptoradar scan-multi "legacy-core:demo_repo/legacy-core-java" "api-layer:demo_repo/api-layer-node" "mobile-adapter:demo_repo/mobile-adapter-python" "pqc-pilot:demo_repo/pqc-pilot-service" -o demo_report.html
MIT licensed. Feedback on the detection approach, and any false positives or false negatives you find, genuinely welcome — link to the repo below.