test(sip): container-restart REGISTER acceptance harness + full FS regression (M6 CP5, B4)

[ryanmurf]

M6 CP5 (B4) — container-restart REGISTER acceptance + full FS regression

Two parts: (1) build the outstanding CP5/B4 acceptance — an on-demand,
isolated-Docker harness that proves dynamic authenticated REGISTER routing
follows a real container restart to a new IP, receiver-side; (2) run the full
isolated FreeSWITCH harness to confirm M6 regressed nothing.

PR #154 landed only the in-process mechanism proof
(crates/asterisk-integration-tests/tests/e2e_dynamic_register.rs). Its literal
plan acceptance — a bridge container restarted so it comes back with a new
container IP on an isolated Docker network, then a call routing to the new IP —
was not built. This adds it as tests/cp5-register-restart/.

Part 1 — the harness (tests/cp5-register-restart/run.sh)

Isolated --internal network. rustisk (fixed IP, kept out of the dynamic pool
via --ip-range) + a sip-bridge container. A and B are read at runtime
(docker inspect), never hardcoded. Every INVITE is verified receiver-side
(the datagram's arrival on the bridge/sentinel), never a rustisk TX log.

• GREEN-A — bridge digest-REGISTERs (401 → digest → 200) advertising IP
  A; Dial(PJSIP/bridge) INVITE arrives at A (captured on the bridge).
• restart — the bridge container is destroyed; a sentinel container
  seizes the vacated A (so a stale-route INVITE has a receiver and Docker
  can't hand A back); a fresh bridge comes up with a new IP B (B ≠ A,
  enforced) and re-REGISTERs from B.
• GREEN-B — Dial(PJSIP/bridge) INVITE follows to B (captured on the
  new bridge) and not to stale A (sentinel silent). rustisk holds both
  the stale-A and fresh-B bindings live, so Registrar::best_contact's
  newest-wins is genuinely load-bearing — not a trivial single-binding route.
• RED (captured) — routing is defeated by a static contact pinned to A
  (bridge_pinned, whose AoR is contact = sip:bridge@A:5060 and never
  re-registers). Dial(PJSIP/bridge_pinned) misroutes to stale A, the sentinel
  catches the INVITE, and the follow-to-B assertion goes RED. This negative
  control proves the A-detection is real (the harness can fail). The
  complementary best_contact oldest-wins RED is covered in-process by
  e2e_dynamic_register.rs.

Captured proof (representative run, A/B assigned at runtime)

bridge IP A (pre-restart, runtime-read): 10.252.68.33
bridge IP B (post-restart, runtime-read): 10.252.68.34
A != B: yes

GREEN-A (INVITE arrives at registered A):        PASS
GREEN-B (INVITE follows to B, stale A silent):   PASS
RED     (static-pin defeats routing -> A caught): PASS_AS_RED

--- bridge receiver-side INVITE captures ---
INVITE role=bridge own=10.252.68.33 src=10.252.68.2:15060 ruri=sip:bridge@10.252.68.33:5060 ...
INVITE role=bridge own=10.252.68.34 src=10.252.68.2:15060 ruri=sip:bridge@10.252.68.34:5060 ...
--- sentinel (stale-A) INVITE captures ---
STRAY_INVITE role=sentinel own=10.252.68.33 src=10.252.68.2:15060 ruri=sip:bridge@10.252.68.33:5060 ...   # RED
--- rustisk registrar binding events ---
INFO Contact registered aor=bridge contact=sip:bridge@10.252.68.33:5060 expires=3600   # stale A
INFO Handled REGISTER status=200
INFO Contact registered aor=bridge contact=sip:bridge@10.252.68.34:5060 expires=3600   # fresh B (both live)
INFO Handled REGISTER status=200

The INVITE follows A → B receiver-side; the RED (static pin) misroutes to
stale A and is caught. This is test infrastructure only — no product code
changes — so merging it is safe.

Part 2 — full isolated FreeSWITCH regression (master @ 8a3cc2d)

tests/freeswitch-pin-gate/run.sh (full case), on the same rustisk build. All
sub-cases PASS — no M6 regression.

Area	Result
Startup fail-closed (absent + invalid mounted secret)	PASS / PASS
PIN gate REJECTED (REJECTED_B_SIDE delta=0; TX/RX voice + DTMF)	PASS
PIN gate GRANTED (GRANTED_BRIDGE M2 two-way; TX/RX voice + DTMF)	PASS
M3 no-input absolute deadline	PASS
Bundle-A AMI positive control (Newexten/PinGate + VarSet/PINGATESTATUS)	PASS
SIP-only pcap	PASS
M1 listen-only (RTPVoiceFramesRx=0), Dial-timeout CANCEL	PASS / PASS
M2 two-way + DTMF A→B, ingress hygiene (wrong-source/wrong-pt/malformed/unstable-SSRC), pattern injection, BYE-silent + deadline-silent teardown	PASS
M4 dropped-200, late-replay, dropped-ACK Timer-H, dropped-BYE, 180-silence Timer-B, forged-dialog, concurrency-10, abandon-200 race, BYE-final-481, BYE-final-dropped Timer-F	PASS (all)
M5 500-call exact-baseline soak (every registry restored to 0…)	PASS
Bundle-B ZERO_HIT_SECRET_AUDIT (both patterns, concatenated + per-digit; every artifact; AMI live)	PASS

Constraints honored

• Isolated Docker only — no touch to the live voice stack / Helm / k8s / trunk /
  Chime / Mumble. Not wired into CI (on-demand).
• Every container/network reaped on exit (host-PID kill for tron's
  userns-remapped docker, where docker rm -f/stop hang).
• Throwaway random test PIN, mounted read-only, shredded on exit; the real PIN
  never appears in code, logs, or evidence.
• cargo clippy --workspace --exclude pjsip-shim -- -D warnings and
  cargo test --workspace --exclude pjsip-shim both green (no Rust changed).

[ryanmurf]

External adversarial review: codex (gpt-5.5, xhigh). It confirmed the receiver-side A/B captures are genuinely distinct, B != A is enforced, and the RFC2617 qop=auth digest matches rustisk's verifier — and flagged 5 false-green-hardening issues, all applied in 681fac6:

1. (P1) sip_agent.py accepted an unauthenticated-200 REGISTER as success → now a hard failure (an auth regression can no longer green-light the digest-auth proof).
2. (P1) an invalid CP5_CASE silently skipped the case-gated assertions and exited 0 → now validated up front.
3. (P2) RED could consume a stray bridge datagram → bridge_pinned now uses a distinct sip:pinned@A contact, and RED correlates on that Request-URI.
4. (P2) reap could leak on a transiently-empty host PID → retry PID acquisition, timeout-wrap every docker op, loud leak alarm.
5. (P2) AMI Originate matched Login's Success → now requires the Originate-specific successfully queued.

Harness re-run green after the fixes (GREEN-A/GREEN-B pass, RED captured); Docker reaped.