docs(reviews): file vision sanity-check from Claude

Reviewer: Claude (Opus 4.8). Filing the strategy/vision-level critique
(for the slice-1 code-review see the companion
2026-06-28-slice-1-claude-adversarial-assessment.md already committed
in 8680f2b).

Filed but not addressed. Six strategic findings:
- S1: wedge leads with determinism (its most contestable leg)
- S2: differentiation valley between spearhead-end and capability ladder
       (the load-bearing critique)
- S3: compliance paradox — regulated buyers won't move on bus-factor-1
- S4: turn-detection ownership at the tap — OpenAI Realtime has its own
       server-side VAD by default; ARCHITECTURE.md says core-authoritative.
       Belongs in the slice-3 (brain) spec as an explicit decision.
- S5: solo scope vs. category timing window
- S6: self-improvement loop is the actual moat, under-sold

Addressing these is strategy-doc work that warrants user signoff and
architecture pre-reading per AGENTS.md; not auto-applied alongside the
code-track fixes.

docs/reviews/2026-06-28-vision-sanity-check.md

@@ -0,0 +1,141 @@
+# Claude Adversarial Sanity-Check — Rutster Vision & Strategy
+
+- **Reviewer:** Claude (Opus 4.8), peer-level read of the strategy docs (not the code — see the companion `2026-06-28-slice-1-claude-adversarial-assessment.md` for the code).
+- **Targets:** `docs/superpowers/specs/2026-06-26-vision-revision-design.md`, `docs/ARCHITECTURE.md`, ADRs 0002–0006, as of `main@22d3f03`.
+- **Question asked:** "Sanity-check the *idea*." This is a strategy critique, not an architecture-correctness review.
+
+## How to use this document (for the GLM-5.2 reviewers)
+
+Same contract as the code assessment: **these are claims to refute, not conclusions to accept.** Strategy has no unit tests, so each critique below carries three fields instead of a falsification test:
+
+- **Steelman (the doc's defense):** the strongest version of the current position. Try to make it win.
+- **What would change the verdict:** the evidence or decision that flips the critique.
+- **Decision implied:** the concrete thing to change *if* the critique holds — so this stays actionable, not philosophical.
+
+The author already pressure-tested this vision once (the 2026-06-26 revision reversed ADR-0001). Hold it to that standard: a critique you can dissolve should be struck with reasoning recorded.
+
+---
+
+## Verdict
+
+**Stronger than "decent."** This is an unusually coherent strategy precisely because it already cannibalized its own founding premises — reversing the Kamailio/rtpengine C shield instead of rationalizing it is the tell that the thinking is honest. The **identity, architecture, and security posture are sound.** The vulnerability is **not the bet** — it's **sequencing and emphasis**: the pitch leads with the one leg a funded competitor can contest, and the moat-grade capabilities sit on the far side of a long solo valley. Both are fixable by reordering, not by changing the thesis.
+
+Confidence: high on the architecture/security being sound; medium on the strategic critiques (they're judgment calls about a market in motion, which is exactly why they're written to be argued with).
+
+---
+
+## TL;DR — strategic findings
+
+| # | Type | One-line |
+|---|------|----------|
+| S1 | **Emphasis** | The technical wedge leads with no-GC determinism — the weakest, most benchmarkable leg — on ground LiveKit Agents will contest directly |
+| S2 | **Sequencing (dominant risk)** | The spearhead (steps 1–6) lands at Vapi/Retell parity; every *differentiating* capability is past it, deep in the capability ladder — a long solo valley before the moat exists |
+| S3 | **Positioning** | The wedge is "strongest for regulated" but regulated buyers are the most allergic to bus-factor-1 infra; first adopters are the builder persona, not the compliance prize |
+| S4 | **Integration risk** | Core-authoritative VAD vs. the brain's own server-side turn detection (OpenAI Realtime) — "AudioOut advisory" meets a vendor with opinions |
+| S5 | **Scope/timing** | Solo attack on a 1.18M-LOC-category problem against a window that may be closing; the long pole is integration/hardening, which AI assistance collapses least |
+| S6 | **Reframe** | The self-improvement loop (own the call → auto-labeled takeover data) is the durable moat and is buried under determinism in the wedge framing |
+
+**Net:** S2 is the one that matters. S1 and S6 are the same fix from two directions (move emphasis off determinism, onto the boundary + data loop). S3–S5 are clear-eyes caveats to record, not blockers.
+
+---
+
+## S1 — The technical wedge leads with its weakest leg
+
+**Type: Emphasis. Confidence: medium-high.**
+
+§3 frames the technical wedge as **(1) no-GC real-time determinism, (2) one secure auditable boundary, (3) operational simplicity**, and calls determinism *"the one thing a Go SFU + external agent structurally can't match on quality."* That ordering puts the most contestable claim first:
+
+- The dominant latency is the **brain round-trip** (hundreds of ms to a speech-to-speech API) — which rutster does not own. No-GC buys **tail smoothness** in the local loop, real but second-order to the RTT.
+- The latency-critical reflex is **local VAD barge-in** — and competitors can also run local VAD. Barge-in *quality* in a no-GC loop is an edge, but it's a number a competitor can benchmark against, not a category they're locked out of.
+- **LiveKit Agents** is the specific threat. The doc characterizes LiveKit as "horizontal media infra, no contact-center domain" — accurate for LiveKit *core* — but LiveKit is actively building the agent/turn-taking layer. Leading on determinism points the pitch at exactly the ground they're moving onto.
+
+The **structural** legs — single auditable boundary (#2) and data-ownership — can't be benchmarked away. Those should lead.
+
+- **Steelman:** for tight turn-taking the floor latency *does* matter, and a no-GC loop genuinely gives more predictable barge-in under load; "structurally can't match" is about the **whole** (self-host + audit + reflex), not the reflex alone — §3's honest caveat already says this.
+- **What would change the verdict:** evidence that no-GC barge-in is a *primary* buying criterion (not a refinement) for the target persona — e.g., a head-to-head where GC jitter visibly degrades a competitor's turn-taking in normal use, not a synthetic stress test.
+- **Decision implied:** reorder §3 — lead the technical wedge with the auditable boundary + data-ownership; keep determinism as the third support, framed as "quality refinement," not "the thing they can't match."
+
+## S2 — The differentiation valley (dominant strategic risk)
+
+**Type: Sequencing. Confidence: high that the valley exists; medium on its severity.**
+
+Walk the spearhead (§10) to its end: *"I called my Rust box, an AI answered, it barged in, and it can't blow my budget."* Motivating and real — but at that milestone the product is **at Vapi/Retell feature-parity while behind them on convenience** (they're managed; this is DIY self-host). Everything *categorically* different is **past** the spearhead, in the capability ladder (§11): human escalation (rung 2), containment measurement (rung 3), the self-improvement loop (rung 4), plus the actual contact-center domain library (ACD/queues/CDR/supervisor).
+
+So a solo builder spends enormous effort to reach "another AI voice bot you happen to host yourself" **before** reaching anything that is a moat. Motivation, validation, and early adoption all have to survive that crossing.
+
+- **Steelman:** the spearhead's job is to prove the *combination is technically real*, not to be differentiated — and "I called my Rust box and an AI answered" is the momentum fuel a solo build needs. Differentiation can follow once the core is proven.
+- **What would change the verdict:** if "self-hostable AI voice bot" is *itself* enough of a draw for the §2 persona (data-ownership alone sells it), the valley is shallow and this de-rates to a caveat.
+- **Decision implied:** pull **rung-2 human takeover** forward to sit adjacent to the spearhead. "AI answers, and the instant it flounders a human seamlessly barges in via the audiohook" is the earliest demo that is **not** Vapi-parity and proves the contact-center thesis — and §11 already notes rung 2 reuses the audiohook primitive PORT_PLAN kept. This reorders the proof so differentiation appears before momentum runs out. (Tradeoff: it widens the spearhead — weigh against "brutally thin.")
+
+## S3 — The compliance paradox
+
+**Type: Positioning. Confidence: medium-high.**
+
+§3 says the wedge is "strongest for regulated/high-trust (PCI/HIPAA/TCPA)." True for *fit* — but regulated buyers are the **most** risk-averse about brand-new, single-maintainer infra in the call path. Self-host cuts both ways: it removes "my data on a vendor's cloud" (good) and adds "we now operate unproven Rust telephony ourselves with bus-factor 1" (bad, for exactly a compliance-conscious org). The realistic first adopters are the **§2 builders** (doing it for a less-regulated client, or internal/experimental use), who *reach* the compliance segment later — the same arc Asterisk walked.
+
+- **Steelman:** self-hosting + auditability + data-ownership are *literally the buying criteria* in regulated spaces, and a builder-led bottom-up motion is how open infra always enters the enterprise — this is a feature of the path, not a bug.
+- **What would change the verdict:** a regulated design partner willing to route real (or even shadow) traffic early — that would prove the compliance segment is reachable before the project is "proven," collapsing the paradox.
+- **Decision implied:** state explicitly that **regulated is the eventual prize reached via the builder persona, not the beachhead.** Keep the §2 builder as the actual day-1 customer in all positioning, so the messaging doesn't chase a buyer who won't move first.
+
+## S4 — Turn-taking ownership at the tap
+
+**Type: Integration risk. Confidence: medium (depends on brain vendor).**
+
+ARCHITECTURE.md is right that the core should be authoritative on VAD/barge-in ("the tap carries the *results* of reflexes, not the responsibility"; "`AudioOut` advisory / core-authoritative"). But the most likely first brain — **OpenAI Realtime** — does **its own** server-side VAD and turn detection by default. Integrating step 3–4 means either:
+- disable the brain's turn detection and feed it clean, locally-detected turns (preserves the reflex-authoritative principle, but is *more* integration work and fights the API's grain), or
+- accept split-brain turn-taking where local VAD and the brain's VAD can disagree (double-triggers, dropped barge-ins).
+
+This is precisely where the elegant "core disposes" principle meets a vendor with opinions, and it's a *design* decision, not an implementation detail.
+
+- **Steelman:** the tap is deliberately core-as-client and format-canonical specifically so the core can drive turns and treat the brain as a dumb speech-to-speech transducer — Realtime supports disabling server VAD, so the clean path is available.
+- **What would change the verdict:** a step-3 prototype showing Realtime-with-server-VAD-disabled gives acceptable turn quality fed by local VAD — then it's just integration, not a tension.
+- **Decision implied:** make "who owns turn detection" an explicit decision in the step-3 (brain) design doc, defaulting to **core-authoritative, brain VAD off**, with the integration cost budgeted. Don't let it be discovered at wiring time.
+
+## S5 — Solo scope vs. category timing
+
+**Type: Scope/timing. Confidence: medium.**
+
+The thin-slice sequencing and AI-pair leverage are the right mitigations and they're real. The honest framing: this is a **1.18M-LOC-category** problem attacked solo, and the long pole isn't writing code (AI helps most there) — it's the **integration/hardening surface** of real-time telephony: NAT traversal, carrier signaling quirks (re-INVITE for hold/transfer, DTMF transport negotiation, PRACK/early media, IP-auth vs registration trunks, SRTP keying on the trunk side), codec edge cases, behavior under load, and security-hardening every wire parser. AI assistance collapses *that* the least. Meanwhile the "AI-era contact center is white space" window is open **now** but contested (incumbents bolting on AI; Vapi/Retell maturing and funded).
+
+- **Steelman:** "engine not product" bounds the scope hard, the capability ladder is genuinely shippable rung-by-rung, and a focused solo builder with AI leverage in 2026 is materially faster than the team-years Asterisk took — the comparison to 1.18M LOC is unfair because most of that is breadth rutster deliberately rejects.
+- **What would change the verdict:** sustained shipping velocity through steps 1–5 (the trunk client especially) on the planned timeline — evidence the hardening pole is shorter than feared.
+- **Decision implied:** ruthlessly sequence toward the **differentiated** demo (see S2), not breadth — every month spent on capability that overlaps Vapi is a month the window can close. Treat the SIP-trunk hardening (ADR-0003) as the schedule risk it is; WebRTC-first ordering already hedges it.
+
+## S6 — Reframe: the self-improvement loop is the actual moat
+
+**Type: Reframe. Confidence: high that it's the most durable advantage.**
+
+The capability-ladder rung 4 — **own the whole call → every human takeover auto-labeled "AI failed here → human did this" → containment analytics + eval sets + fine-tune export, all on the operator's own infra** — is the one advantage that is:
+
+1. **structurally impossible** for cloud bots (they don't own your data) and CCaaS (not their business model),
+2. a **compounding data-network-effect** (better the more you use it), and
+3. **uniquely enabled** by self-hosting + owning the entire call.
+
+That is a durable moat in a way no-GC determinism is not. The north-star one-liner already names it ("a closed ML loop that learns from every escalation") — but the §3 *technical-wedge* framing buries it under determinism. S1 and S6 are the same correction from opposite ends: **the wedge is the auditable boundary + the data loop; determinism and simplicity are supporting cast.**
+
+- **Steelman:** the loop is rung 4 — years out — so leading with it oversells what exists; determinism is what the *spearhead* actually demonstrates, so it's honest to lead with it now.
+- **What would change the verdict:** nothing about *whether* the loop is the moat; the open question is *timing of the claim* — see decision.
+- **Decision implied:** separate "what the spearhead proves" (the combination works) from "what the wedge is" (boundary + data-ownership + loop). Lead strategy/positioning with the latter even while the spearhead demonstrates the former. Don't let the demo's emphasis become the strategy's emphasis.
+
+---
+
+## Architecture decisions — spot-check (mostly sound; recorded for completeness)
+
+These are not strategy critiques; they're the architecture calls the strategy rests on, and they hold up:
+
+- **Remove the control↔media gRPC hot-path hop / fused per-call vertical** (ADR-0002) — correct; that hop was always a latency + failure-mode liability.
+- **Spend gate co-located with trunk termination** — genuinely strong and *under-sold*; "a runaway brain can't exceed spend because it doesn't hold the wire" is a demonstrable structural property a 3-vendor stack cannot match. Promote it in the pitch.
+- **Core-as-client tap, no inbound port; core-authoritative playout** — correct security instinct; deletes an attack class.
+- **WebRTC-first ingress** (ADR-0006) — smart de-risking; first-call never blocks on SIP.
+- **Rust-native trunk SIP, no C shield** (ADR-0003) — right call for the scoped use case (interop tail collapses to a few cooperative carriers; parser-security thesis becomes literally true). Mild caveat: even cooperative carriers carry a provider-independent signaling-state-machine complexity that "a few documented providers" slightly undersells — the cost is accepted, just don't let the framing make it sound small.
+- **Valkey event bus, 20 ms loop never on the bus, bus not source-of-truth for billing-critical state** (ADR-0005) — correct discipline.
+
+---
+
+## Verification checklist for the GLM-5.2 reviewers
+
+1. **S2 is the load-bearing critique.** Decide whether "self-hostable AI voice bot" is independently compelling to the §2 persona. If yes, S2 de-rates and the spearhead stands as-is. If no, pull rung-2 takeover forward. Everything else is secondary to this call.
+2. **S1 + S6 are one decision:** move the wedge's emphasis off determinism onto boundary + data loop — or refute that determinism is benchmarkable/contestable.
+3. **S4 belongs in the step-3 design doc**, not here — make sure it lands there as an explicit turn-ownership decision.
+4. Push back on **S3/S5** with evidence (a regulated design partner; demonstrated velocity) — those are the critiques most likely to dissolve with a single data point.
+5. Strike anything you can dissolve, with reasoning recorded — same standard the vision held itself to when it reversed ADR-0001.