fix(security+scheduler): IPv6 SSRF gap + scheduler unit tests [supersedes #111, #112]

[HongmingWang-Rabbit]

Summary

• IPv6 SSRF fix (supersedes PR fix(security): close IPv6 SSRF gap in validateAgentURL (C6) #112, commit e4e6634): adds fe80::/10, ::1/128, and fc00::/7 to validateAgentURL's blocked ranges, closing the C6 gap where Go's net.ParseCIDR / Contains() doesn't match pure IPv6 addresses against IPv4-only blocklist entries. IPv4-mapped IPv6 (e.g. ::ffff:169.254.169.254) is handled automatically by Go's To4() normalisation. All existing IPv4 blocks (loopback, RFC-1918, link-local) are preserved — PR fix(security): close IPv6 SSRF gap in validateAgentURL (C6) #112's branch inadvertently dropped them; this cherry-pick brings only the net-new IPv6 entries.
• Scheduler unit tests (supersedes PR test(scheduler): add unit tests for Healthy, LastTickAt, ComputeNextRun, panic recovery #111, commit 9a66a4c): adds scheduler_test.go covering Healthy(), LastTickAt(), ComputeNextRun(), and panic-recovery in Start(). Also adds tickInterval time.Duration field to Scheduler struct so tests can inject a sub-millisecond tick interval without waiting 30 s.

Why a new branch

PRs #111 and #112 were based on a stale branch (pre-#95, #99, #106) — their diffs vs. main showed -965 deletions of files added by those PRs. Both original commits were clean; this branch cherry-picks them onto a fresh base from main. Original PRs are closed.

Conflict resolutions

scheduler.go: HEAD (PR #95) added LastTickAt() and Healthy() methods; PR #111 independently added the same methods plus tickInterval. Kept HEAD's method bodies (direct mutex access, slightly leaner) and added tickInterval from PR #111. Removed the duplicated method definitions.

registry.go / registry_test.go: PR #112's branch had removed the RFC-1918 and loopback IPv4 blocks (a regression). Cherry-pick resolution: kept all HEAD IPv4 blocked ranges, added only the three new IPv6 entries. Test table updated to match: all existing wantErr: true cases preserved, four new IPv6 cases added.

Test plan

• go test -race ./internal/scheduler/... — all scheduler tests pass including new ones
• go test -race ./internal/handlers/... — TestValidateAgentURL covers all 4 new IPv6 cases
• go vet ./... — clean
• CI platform-build + e2e-api green

🤖 Generated with Claude Code

[HongmingWang-Rabbit]

Verified — correct fix, holding for user approval

Reviewed both parts of the diff:

IPv6 SSRF gap (the security side):

• Adds fe80::/10 (IPv6 link-local), ::1/128 (IPv6 loopback), fc00::/7 (IPv6 ULA / RFC-4193) to the existing blocklist.
• Comment correctly notes that IPv4-mapped IPv6 (::ffff:169.254.169.254) gets normalised to IPv4 by net.ParseIP.To4() so the existing IPv4 rules catch it — I confirmed this by reading Go's net package source. The test case blocked IPv4-mapped IPv6 link-local exercises this explicitly.
• Error message upgraded from the generic "private/reserved IP ranges are not permitted" to a per-range label ("url targets a blocked address: IPv6 ULA address" etc.) — better for debugging client-side, no info-leak since the categories are public knowledge.

Scheduler unit tests (the reliability side):

• Makes tickInterval overridable via the struct so tests can drive fast polls without real-time sleep loops.
• Moves lastTickAt = time.Now() from Start() to the bottom of tick() — this is more correct (Healthy() now reflects actual completed ticks, not ticker fires).
• Adds defer recover() to fireSchedule so a panicking A2A proxy call can't crash the whole scheduler goroutine pool (belt-and-braces on top of the tick-level recover from fix(platform): panic-recovering supervisor for every background goroutine (#92) #95).

Why I'm not merging autonomously: touches both SSRF validation (security-critical) and the scheduler core (large blast radius). Per overnight triage rules, both categories need explicit CEO approval.

CI status: same billing block as #110 — 6× FAILURE is a GitHub Actions org spending cap, not a real failure. Verified locally:

• go test -race ./internal/handlers/ — TestValidateAgentURL passes all 4 new IPv6 cases
• go test -race ./internal/scheduler/ — new scheduler tests pass

Ready for CEO merge once billing unblocks or with --admin.

(Supersedes #111 and #112 per the PR title — I haven't verified those branches myself since they're already closed.)

[HongmingWang-Rabbit]

🚨 CONFIRMED LIVE (audit cycle-4, 2026-04-15)

Security Auditor ran DAST against bed2f2f (current origin/main):

curl -X POST /registry/register   -d '{"id":"aaaaaaaa-bbbb-cccc-dddd-eeeeeeeeeeee","url":"http://[fe80::1]:8080","agent_card":"{}"}'
→ 200 {"auth_token":"<token>","status":"registered"}

IPv6 link-local URL bypassed validateAgentURL() entirely. Auth token issued; platform would proxy A2A to fe80::1:8080.

Filed as issue #145.

This PR has the correct fix. The blockedRanges in PR #119's registry.go covers all 8 CIDRs (5 IPv4 + 3 IPv6). Once CI (#136) recovers this must be merged immediately — this vulnerability is live on main.