feat(web-security): postMessage listener discovery and origin-validation depth

[GangGreenTemperTatum]

Across all 82 skills, these appeared zero times:

MessagePort · MessageChannel · __lookupSetter__ · nr@original · __sentry_original__ · zone.js

Our skills told an agent to grep addEventListener("message", ...). That silently misses:

• Wrapper-hidden handlers. Sentry/New Relic/Rollbar/Raven/Bugsnag/Zone.js wrap listeners, so the registered function is the monitoring shim — reviewing it reads the wrong code entirely.
• window.onmessage = fn setter assignment.
• MessagePort/MessageChannel handlers (iframe & worker bridges, SDKs).

Adds a wrapper tell → recover-from table verified line-by-line against src/main.ts, plus a runtime capture snippet hooking all three surfaces.

Origin validation depth

Previously we named only startsWith/endsWith. Now the full set — indexOf, includes, search, startsWith, endsWith, loose ==/!= — plus two regex bugs that read as correct:

Bug	Example	Bypass
Unescaped dot	/^https:\/\/trusted.example\.com$/	https://trustedXexample.com
Missing $ anchor	/^https:\/\/trusted\.example\.com/	https://trusted.example.com.evil.tld

Also flags wildcard targetOrigin on the reply path — leaks even when the inbound check is sound.

Two errors caught during review (worth flagging)

1. The documented rg command didn't run. My weak-origin pattern used (?!=) lookahead, which ripgrep's default engine rejects outright (error: look-around ... is not supported). Replaced with a portable [^=] form, and there's now a test that shells out to real rg and asserts it matches the 7 unsafe lines while skipping strict ===.
2. The native code check was overstated. I'd presented addEventListener.toString().includes('native code') as a detection. It's trivially spoofable by overriding toString (demonstrated) and also trips on unrelated extensions. Now qualified as a weak signal, and replaced with a snippet that genuinely enumerates listeners — verified in headless Chromium capturing all three surfaces:

   ["addEventListener:appHandler","onmessage:inlineHandler","MessagePort:portHandler"]
   

Also corrected against upstream source: Rollbar tries _wrapped before _rollbar_wrapped, Raven unwraps via its single function-valued property, and Bugsnag is not recoverable (upstream dropped the deprecated callee.caller chain) — my first draft had all three wrong.

Tests

522 passing (+43). Assertions are behavioural, not textual:

• every documented rg pattern is compiled and executed
• every origin bypass in the table is actually run
• wrapper claims are checked against upstream source
• all 10 rule IDs and the example output verified against the live engine

Mutation-tested. Deleting any of the three new sections fails 6–7 tests. Two guards were tightened after mutations initially slipped through: an operator name matched prose outside its table, and not recoverable appeared twice so corrupting one row still passed.

Test plan

• pytest tests/ → 522 passed
• dreadnode capability validate --strict → 0 failed (only environmental caido-cli/burp checks)
• ruff check --select E9,F63,F7,F82 → clean
• Both skills load through the real Dreadnode skill loader
• Runtime capture snippet verified in headless Chromium
• All ast-grep/rg commands executed against realistic JS