feat(risk-presidio): drop IPv6 short-form and IPv4 unspecified IP false positives

[mfbx9da4]

Why

Production risk_results analysis on the read replica (1.38M total matches) showed Presidio's IP_ADDRESS detector dominating noise — 1.03M of the 1.38M findings, 75% of all triggered risks, come from a single entity type, almost all of them not-actually-IP text fragments.

Breakdown of the noise:

Match pattern	Hits
:: (already filtered)	1,006,890
IPv6 short-form <hex>:: (b::, ad::, bed::, dead::, 1::, 2::, 3::, 5::, d::, deaf::, fad::)	~17,500
0.0.0.0	~640 (across IP_ADDRESS + pii.ip_address)

These come from hex constants, partial address text in logs/code, and 0.0.0.0 bind-address literals. None of them are addresses anyone meaningfully uses.

What

Extend isPresidioFalsePositive to drop:

1. All unspecified addresses via netip.ParseAddr(...).IsUnspecified() — automatically covers ::, ::0, 0::0, 0:0:0:0:0:0:0:0 (existing), 0.0.0.0 (new), and any other spelling Go's parser recognizes.
2. IPv6 short-form <hex>:: via regex ^[0-9a-f]{1,4}::$ — catches the new ~17k class.

::1 (loopback) and dead::beef (two-group IPv6) are deliberately preserved as real addresses — covered by existing + new test cases.

Test plan

• Unit tests in presidio_internal_test.go updated to cover 0.0.0.0, b::, dead::, 1::, DEAF:: (case-insensitive), and the dead::beef / ::1 negative cases
• go vet ./internal/background/activities/risk_analysis/... clean
• go test ./internal/background/activities/risk_analysis/ -run 'IsPresidioFalsePositive|FiltersIPv6Unspecified' passes

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review to trigger a review and subscribe this PR to future pushes, or @claude review once for a one-time review.

_{Tip: disable this comment in your organization's Code Review settings.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
gram-docs-redirect	Ready	Preview, Comment	May 19, 2026 10:52am

[changeset-bot]

🦋 Changeset detected

Latest commit: 0940bde

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
server	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[github-actions]

Prompt-injection risk report

Corpus: 933 cases (384 malicious / 549 benign)
This PR: 9331fe91 · 2026-05-19T10:53:50Z

No main baseline artifact found yet; this comment shows the current run only.

Operational Modes

mode	status	total	TP	FP	FN	precision	recall	F1	FP-rate
L0 only	ok	933	76	0	308	1	0.1979	0.3304	0
L0 + L1 opt-in	skipped	-	-	-	-	-	-	-	-

L1 opt-in was not evaluated in this run: classifier URL is not set.

_{Generated by .github/scripts/risk-metrics-comment.py. Full source/rule breakdown and samples are in the risk-accuracy-metrics artifact.}

[speakeasybot]

🚀 Preview Environment (PR #2915)

Preview URL: https://pr-2915.dev.getgram.ai

Component	Status	Details	Updated (UTC)
✅ Database	Ready	Existing database reused	2026-05-19 10:58:15.
✅ Images	Available	Container images ready	2026-05-19 10:57:59.

_{Gram Preview Bot}