Pr3 extractors ir

[AtharvKshirsagar]

What

Adds signals to the CanonicalIR that the existing extractors structurally could not see.
No new IR — same model, new node/edge types so the facts exist to diff.

IAMPostureExtractor (new)

IAMEdgeExtractor only emits CAN_DO when a statement's Resource resolves to an
in-stack Ref. Three whole classes of privilege were therefore invisible:

Signal	Was	Now
AWS managed policies (AdministratorAccess, PowerUserAccess)	no edge at all	expanded to real permissions
Resource: "*"	no edge at all	edge to synthetic Resource::Wildcard
sts:AssumeRole trust policies	not read	cross-principal edges

Standalone roles (a CI/deployment role with no compute attached) get a synthetic
principal node — a role holding AdministratorAccess is an escalation regardless
of what assumes it.

PublicEntryExtractor (new)

A Lambda Function URL is a public HTTPS endpoint on the function itself — no security
group, no load balancer. Invisible to SG-based reachability. Now emits an
Internet → function NETWORK_PATH edge.

AuthControlExtractor (extended)

Handles AWS::Lambda::Url — AuthType: AWS_IAM vs NONE.

Precision guardrails

Two filters prevent this from becoming a false-positive generator:

• Telemetry ≠ privilege. CDK gives every Lambda xray:PutTraceSegments on
  Resource: "*". Wildcard statements on telemetry/read-only services are excluded —
  without this, precision dropped to 46%.
• Benign execution roles. AWSLambdaBasicExecutionRole and friends are explicitly
  listed as non-escalating.

Impact

Privilege drift detection 1/10 → 6/10 on the benchmark with detector logic
essentially unchanged — evidence that extraction, not detection, was the bottleneck.
Combined with #: recall 20% → 30%, precision 59% → 94%.

Tests

146 passing. One pre-existing test_graphdiff.py failure on drift-classes, unrelated.