PRD: AI Platform Maturity & Production Readiness

Document ID: PRD-004 File Path: docs/01-product/prd-ai-platform-maturity.md Version: 1.0.1 Status: In delivery — Phases 1–2 of the derived ticket docs are complete and Phase 3 is partial (22 of 39 tickets done; see the v1.1 index); Phase 3’s remaining ecosystem items are re-prioritized through PRD-005’s lens per the v1.2 milestone Owner: Product / AI Engineering Last Updated: 2026-07-15

---

1. Purpose

PRD-003 closed the deployed-vs-designed gap that made the GA appliance defensible: authentication, crash-safety, an execution driver, a contract-tested API. This PRD closes the next gap — the capability-vs-credible-AI-product gap — surfaced by a 2026-07-09 full-project engineering audit (five parallel deep-dives: AI core, workflow+server, dashboard UI, tools/plugins/sandbox, and DX/CI/deployment/docs).

Where PRD-003 asked “is the platform safe and honest?”, this PRD asks “is it a capable, operable AI product a team would actually build on?” The audit’s one-sentence verdict:

Wovyr has excellent primitives, but three structural problems recur at every layer: sophisticated machinery is built and tested yet never wired into the real run path; the AI core is missing table-stakes production capabilities; and single-node assumptions are baked into surfaces the product markets as multi-tenant.

This document turns the audit’s ~90 findings into numbered workstreams and testable requirements, then hands them to three phased ticket docs (RM-AIM-P1, RM-AIM-P2, RM-AIM-P3).

This is a planning input, not a promise. Requirements graduate to committed work through the roadmap (v1.1) and, where they change a boundary contract, through an ADR.

---

2. Problem Statement

The audit confirmed PRD-003’s hardening shipped and holds. The remaining gaps are not safety gaps — they are capability, correctness, and operability gaps:

1. The AI core is a thin loop. run_agent has no context-window management (the full history is cloned into every request and grows unbounded), no tokenizer, sequential-only tool execution, and no native Anthropic provider. Worst of all, cost tracking is hardcoded to $0 for every real provider, which silently disables the per-project quota enforcement PRD-003 built.

2. The strong machinery is dead code. Container/gVisor/Firecracker sandboxes, the warm SandboxPool, and the FairScheduler are all built and tested — but the run path hardcodes native_only(), so none of them ever run. On Windows the “sandbox” enforces only a timeout. The distributed workflow/queue/lease layer has the same shape (PRD-003 Path A deferred it, correctly).

3. The RAG stack is missing its middle. No chunking (whole documents get one diluted embedding), no re-ranking, a semantic cache whose key omits the system prompt and embedding-model id (→ wrong-context hits), and no retrieval-quality evaluation.

4. The evaluation harness is a prototype, not a gate. Substring/exact scoring only; no LLM-as-judge, no baselines, no thresholds, no variance measurement.

5. The ecosystem has no connective tissue. No MCP (Model Context Protocol) client to reach external tool servers, no plugin authoring SDK/scaffold, a tiny read-biased builtin toolset (no fs_write, code-exec, or web-search).

6. Operability and DX lag the code. Version pinned at 0.1.0 despite a v0.3.0 tag, no CHANGELOG, no release automation, no published container image; the dashboard and Windows are both absent from CI; the CLI reference doc misdescribes the primary entry point; the dashboard has zero tests and stores bearer tokens in localStorage.

The encouraging counterweight, again: because the abstractions are correct, most fixes are “wire the good primitive onto the path” or “add the missing stage,” not “redesign.”

---

3. Baseline (as of 2026-07-09)

• Shipped & hardened (v0.1–v1.0 GA): agent runtime, workflow engine, memory engine, LLM gateway, tool runtime, plugin engine + marketplace, multi-tenancy, events/webhooks, audit, secrets, KMS, the RM-GA-P1..P4 security/durability/ contract floor, and (2026-07-09) the GA-003 residual findings + an S3-compatible backup destination.
• The gap this PRD closes: the difference between a hardened appliance and a capable, operable, extensible AI product — the AI-core depth, RAG maturity, eval rigor, ecosystem connectivity, sandbox activation, distributed correctness, and DX/release/UI hygiene that the audit found missing.

---

4. Goals & Non-Goals

4.1 Goals

• Make the platform’s own production claims true: real cost accounting (so quotas mean something), context/token management (so long runs don’t silently truncate), activated sandboxing (so isolation is real), and durable/graceful server behavior (so a restart isn’t data loss).
• Make it a credible AI product: native Anthropic support, structured output, guardrails, a real RAG middle (chunking + reranking), and an evaluation gate that can actually catch a regression.
• Make it extensible: MCP, a plugin authoring SDK, and a richer, ergonomic tool surface.
• Make it operable and adoptable: consistent versioning + release automation + a published image, the dashboard and Windows in CI, accurate docs, and a UI with tests and a shared component system.

4.2 Non-Goals

• Billions-of-memories / thousands-of-concurrent-runs capacity engineering at real scale (PRD-002 post-GA Scale work).
• Marketplace monetization / billing (PRD-002).
• Multi-region residency (PRD-002).
• A cloud-KMS/HSM-backed root key (fast-follow; tracked in GA-003).

---

5. Requirement Conventions

• Requirements are R-<workstream>.<n>, testable, and cite the audit finding(s) they close (§12 traceability). Each maps to one or more tickets in the phase docs.
• Priority: P0 = correctness/data-loss/quota-integrity blocker, P1 = needed for a credible product, P2 = quality bar, P3 = fast-follow.
• Severity below mirrors the audit’s own (High/Med/Low).
• “Done” = implemented, covered by an automated test that fails on regression, and the linked docs/ spec updated (and for wire changes, both SDKs + openapi.yaml).

---

6. Workstreams

Twelve workstreams (A–L). Each requirement’s ticket code and phase are in §12.

WS-A — AI Core Runtime

• R-A.1 (P0) — Context-window management: add a tokenizer and a token-budgeted history compactor (drop/summarize oldest tool turns) before each model call; stop cloning unbounded history. (Audit: no context mgmt, no token counting.)
• R-A.2 (P0) — Execute independent tool calls in one turn concurrently, ordered by call id. (Sequential tool loop.)
• R-A.3 (P1) — Apply the agent manifest’s max_steps as the default budget in run_agent_inner (not only via wovyr-runtime). (max_steps ignored.)
• R-A.4 (P1) — Step-level error recovery: retry a recoverable model-step error; on budget exhaustion, force a final tool-less answer instead of hard-erroring.
• R-A.5 (P2) — Richer streaming: emit tool-call-argument and reasoning/thinking events, not just content deltas.

WS-B — Provider & Cost

• R-B.1 (P0) — Per-model price table; compute cost_usd from returned token usage in every provider. (Cost hardcoded $0 → quotas are a no-op.)
• R-B.2 (P1) — First-class AnthropicProvider (Messages API: native tool-use, system handling, prompt caching, extended thinking). (No native Claude.)
• R-B.3 (P1) — Add response_format/tool_choice/json_schema to ChatRequest and translate per provider (JSON mode, forced tool). (No structured output.)
• R-B.4 (P2) — Normalize/validate tool JSON-schema; surface tool-arg parse failures back to the model instead of passing null. (Verbatim pass-through.)
• R-B.5 (P2) — Multimodal content parts (image/audio) on Message.content.
• R-B.6 (P3) — Retry jitter + honor Retry-After; distinguish 429 vs 5xx.

WS-C — Memory & RAG

• R-C.1 (P1) — Document chunking with parent-document linkage before embedding.
• R-C.2 (P1) — A re-ranking stage (cross-encoder or LLM reranker) after RRF.
• R-C.3 (P1) — Fix the semantic-cache key (include system prompt + tools) and stamp/verify the embedding-model id on every entry. (Wrong-context hits.)
• R-C.4 (P2) — BM25/TF-IDF in-process keyword search for backend parity.
• R-C.5 (P2) — Real timestamps on records + range/time metadata filters.
• R-C.6 (P3) — Incremental re-embedding / embedding-model migration.

WS-D — Evaluation

• R-D.1 (P1) — LLM-as-judge + semantic-similarity scoring alongside substring.
• R-D.2 (P1) — Turn wovyr-eval into a real gate: golden baselines, pass-rate thresholds, variance-over-N, persisted score artifacts compared in CI.
• R-D.3 (P2) — Evaluate the RAG path (run_agent_with_memory) and manifest max_steps; add retrieval metrics (recall@k / nDCG / MRR).

WS-E — Sandbox & Tool Runtime

• R-E.1 (P0) — Wire SandboxManager::detect() + SandboxPool into the agent/ server run path so container/gVisor/Firecracker actually run. (Dead code.)
• R-E.2 (P0) — Windows Job Object for memory/CPU/PID limits in the non-Unix native path. (Windows = timeout only.)
• R-E.3 (P1) — Confined fs_write builtin; R-E.4 (P1) — a sandboxed code-execution tool; R-E.5 (P2) — a #[derive(Tool)]/schemars schema+typed- param ergonomics upgrade.
• R-E.6 (P2) — Document the platform matrix and fail closed when egress lockdown is unavailable (non-Linux).

WS-F — Ecosystem & Extensibility

• R-F.1 (P1) — An MCP client tool-source (stdio/HTTP) proxying external tools into ToolRegistry. (No external tool servers at all.)
• R-F.2 (P1) — A plugin authoring SDK crate + wovyr plugin new scaffold (manifest + wasm build + digest computation + trust snippet). (Format docs only.)
• R-F.3 (P2) — A container capability loader (reuse ContainerSandbox).
• R-F.4 (P2) — One-shot wovyr plugin publish (sign + fill digests + emit trust).
• R-F.5 (P3) — Marketplace OSV/CVE feed keyed on SBOM name@version.

WS-G — Server & Multi-Tenancy

• R-G.1 (P0) — Graceful shutdown/drain (with_graceful_shutdown + SIGTERM).
• R-G.2 (P1) — Durable async-run store (or documented non-durability), so a restart doesn’t orphan pollable runs.
• R-G.3 (P1) — Durable webhook outbox + delivery worker with persisted DLQ.
• R-G.4 (P1) — API-key lifecycle: created/expires/revoked metadata, a revoke endpoint, rotation, last-used. (Mint-only today.)
• R-G.5 (P1) — Distributed rate limiting (shared store) for multi-node.
• R-G.6 (P1) — Per-tenant token quotas; enforce or remove the two dead quota dimensions; per-tenant rate tier.
• R-G.7 (P2) — Tenant-configurable daily-cost reset boundary (timezone).
• R-G.8 (P2) — Cache FileApiKeyStore in memory; served OpenAPI; request-path unwrap audit; idempotency-store write-amplification fix; move the ~2,260-line inline lib.rs test suite out.

WS-H — Workflow Engine

• R-H.1 (P0) — Postgres connection pool (deadpool/bb8) + reconnect/health.
• R-H.2 (P0) — Sub-workflow recursion depth guard / ancestor-cycle detection.
• R-H.3 (P1) — TLS to Postgres; R-H.4 (P1) — fenced event-sequence generation (DB identity/sequence + lease-token fencing), replacing MAX(seq)+1.
• R-H.5 (P1) — Loop / for-each (map-over-collection) activity.
• R-H.6 (P1) — Dynamic (data-driven) fan-out.
• R-H.7 (P2) — Checkpoint size cap + out-of-line large activity outputs; event- log compaction + paged load; indexed list() columns + SQL-side pagination; fire_at-indexed timers + adaptive dispatch sleep.
• R-H.8 (P3) — Activity progress events; event-enum schema versioning.

WS-I — Guardrails & Prompt Management

• R-I.1 (P1) — Content-safety / moderation / PII-redaction hooks on model input and output in the agent loop.
• R-I.2 (P2) — A prompt template/versioning registry (variables, versions, A/B).
• R-I.3 (P1) — Default the secrets store to encrypted-at-rest (plaintext becomes the explicit opt-out); R-I.4 (P2) — audit-log time-range + cursor pagination + indexed sink; R-I.5 (P3) — request-scoped secret channel (vsock/stdin) instead of WOVYR_SECRET_* env injection.

WS-J — DX, SDKs & Release

• R-J.1 (P1) — Reconcile versioning: bump workspace/badges/SDKs to the real tag, add a maintained root CHANGELOG.
• R-J.2 (P1) — Release automation: tag-triggered signed binaries + a published container image (GHCR/Docker Hub) + npm/PyPI SDK publish + generated changelog.
• R-J.3 (P1) — Add the dashboard (build/lint/test) and a Windows matrix leg to CI.
• R-J.4 (P2) — SDK parity: async Python client, mutation retry (with Idempotency-Key), a wait_for_completion poll helper, TS paginateAll, coverage/benchmark tracking in CI.
• R-J.5 (P2) — SDK versioning tied to API version + per-SDK CHANGELOG + server/ SDK skew warning; reconcile the Python PyPI-publish claim.
• R-J.6 (P2) — Regenerate docs/11-cli/commands.md from the real clap tree; add per-doc shipped/aspirational status front-matter; add a top-of-README 5-minute quickstart; unify the repository URL across manifests.
• R-J.7 (P3) — Decide/document Go/Java clients (roadmap or non-goal).

WS-K — Dashboard / UI

• R-K.1 (P1) — Move bearer token off localStorage (in-memory/session or BFF-issued httpOnly cookie). (XSS-exfiltratable.)
• R-K.2 (P1) — UI test coverage: service specs (SSE parser, manifest round-trip)
  • a smoke e2e; drop the global skipTests:true.
• R-K.3 (P2) — Shared component library (StatusPill/Tabs/Modal/Table/empty/ loading/error), replacing the duplicated statusClass/errText patterns and the native confirm().
• R-K.4 (P2) — Share API types with sdks/typescript (or generate from OpenAPI); replace string-built YAML with a real (de)serializer; central HTTP error handling (no swallowed errors).
• R-K.5 (P2) — Audit-log viewer surface; R-K.6 (P2) — responsive/mobile breakpoints; R-K.7 (P2) — accessibility pass (label associations, aria on icon buttons, modal focus mgmt).
• R-K.8 (P3) — Prompt playground; live nav badges (or remove the fakes); i18n decision; icon sprite for bundle hygiene.

WS-L — Observability & Operability

• R-L.1 (P1) — Per-tenant/per-project metric labels (bounded cardinality).
• R-L.2 (P2) — Queue-depth / in-flight / pending-timer / webhook-DLQ gauges.
• R-L.3 (P2) — Spans around Postgres/queue/dispatcher operations.
• R-L.4 (P2) — A systemd unit + install script for the single-node appliance; an operator upgrade/backup/migration runbook.
• R-L.5 (P3) — SLO / error-budget burn-rate metrics + starter alert rules; Helm HA/TLS templating; a minimal Terraform module (or explicit scope-out).

---

7. Distributed Scale-Out (folded from PRD-003 Path B)

PRD-003 deferred the distributed platform (Path B) to a “v1.1 Scale-Out” milestone. Several of its wiring tickets overlap this PRD’s correctness work and are absorbed here as P1 items where they are single-node-correctness bugs today (R-G.5 distributed rate limiting, R-H.1 pooling, R-H.4 fencing), and left as PRD-002 capacity work where they are pure scale. The multi-replica shared-catalog promotion (PRD-003 R-5.1/R-5.2) remains gated on Product demand and is tracked, not committed, in v1.1 §Scale-Out.

---

8. Prioritization & Sequencing

Three phases, ordered by dependency (calendar dates omitted per house convention). Within a phase, items are parallelizable.

• Phase 1 — Make production claims true (P0/P1). The fixes that make existing features actually work: R-B.1 (cost table) is the single highest-leverage item — it silently disables quota enforcement everywhere. Plus context/token mgmt, sandbox activation, graceful shutdown, durable async runs, Postgres pool, release reconciliation, and dashboard+Windows CI. → RM-AIM-P1
• Phase 2 — Credible AI product (P1/P2). Anthropic provider, RAG chunking + reranking, semantic-cache correctness, structured output, guardrails, the eval gate, distributed rate limiting, per-tenant quotas. → RM-AIM-P2
• Phase 3 — Ecosystem & scale (P2/P3). MCP, plugin SDK, workflow loops/fan-out, encrypted-secret default, UI component library + audit viewer + responsive, SDK parity, docs, systemd/runbooks, observability gauges. → RM-AIM-P3

The trap to avoid: shipping more surface (or more UI) before R-B.1 lands — every cost/quota number the product reports until then is fiction.

---

9. Exit Criteria

This PRD’s scope is met when:

1. Every real provider reports accurate cost_usd; per-project quotas enforce real spend (WS-B/R-B.1). A long tool loop never silently exceeds the context window (WS-A/R-A.1).
2. The agent run path uses the strongest available sandbox for the run’s trust class; Windows runs enforce real resource limits (WS-E).
3. Claude is a first-class provider; structured output and moderation hooks exist (WS-B, WS-I).
4. Memory chunks + reranks; the semantic cache never serves a wrong-context hit (WS-C). wovyr-eval fails CI on a real quality regression (WS-D).
5. A server restart loses no pollable run or pending webhook; shutdown drains (WS-G). Multi-node rate limits and quotas are correct (WS-G/WS-H).
6. Versioning/CHANGELOG/release automation exist and a container image is published; the dashboard and Windows are in CI (WS-J).
7. MCP and a plugin SDK exist; the UI has tests and a shared component system (WS-F, WS-K).

---

10. Risks & Assumptions

• R: R-B.1 (cost table) changes quota behavior from “always $0, never blocks” to “actually blocks” — a behavior change operators may not expect. Mitigation: ship behind a documented rollout; log computed cost before enforcing.
• R: WS-A context compaction can change agent outputs (summarizing history). Mitigation: make the strategy configurable; default to lossless drop-oldest-tool-turns before summarization.
• R: WS-E activating strong sandboxes changes latency/perf characteristics. Mitigation: the warm SandboxPool already exists to amortize; benchmark first.
• A: The trait-port architecture holds (PRD-003 confirmed), so provider/sandbox/ reranker additions are new impls behind existing traits, not redesigns.
• A: The audit’s file:line evidence (captured 2026-07-09) is current; each ticket re-verifies before implementing.

---

11. Related

• PRD-001 · PRD-002 · PRD-003
• 18-roadmap/v1.1/index.md — the milestone + phase ticket docs
• 18-roadmap/v1.0.md — the GA milestone this builds on

---

12. Traceability Matrix — Findings → Requirements → Tickets

Severity is the audit’s. Phase/ticket codes are authoritative; the phase docs implement exactly these.

Audit finding (abbrev.)	Sev	Req	Ticket	Phase
No context-window mgmt; no tokenizer	High	R-A.1	AIC-101	1
Sequential tool-call execution	High	R-A.2	AIC-102	1
Manifest max_steps ignored by run_agent	Med	R-A.3	AIC-103	1
No step-error recovery / budget-exhaust discards work	Med	R-A.4	AIC-201	2
Streaming is content-only	Low	R-A.5	AIC-202	2
Cost tracking hardcoded $0 (no price table)	High	R-B.1	PRV-101	1
No native Anthropic provider	High	R-B.2	PRV-201	2
No structured output / tool_choice	Med	R-B.3	PRV-202	2
Tool-schema pass-through; null-swallow	Med	R-B.4	PRV-203	2
No multimodal	Med	R-B.5	PRV-204	2
Retry no jitter / ignores Retry-After	Low	R-B.6	PRV-205	2
Memory: no chunking	High	R-C.1	RAG-201	2
Memory: no re-ranking	High	R-C.2	RAG-202	2
Semantic-cache key / embedding-model mismatch	High	R-C.3	RAG-203	2
Naive keyword search (no BM25)	Med	R-C.4	RAG-204	2
No real timestamps / range filters	Low	R-C.5	RAG-205	2
No incremental re-embedding	Low	R-C.6	RAG-301	3
No retrieval-quality eval	Med	R-D.3	EVL-203	2
Eval scoring substring-only (no LLM-judge)	High	R-D.1	EVL-201	2
Eval not a regression gate	High	R-D.2	EVL-202	2
Strong sandboxes dead code (native_only)	High	R-E.1	SBX-101	1
Windows: no resource limits	High	R-E.2	SBX-102	1
No fs_write builtin	Med	R-E.3	SBX-301	3
No code-execution tool	Med	R-E.4	SBX-302	3
Custom-tool ergonomics (no derive)	Med	R-E.5	SBX-303	3
Egress lockdown Linux-only, silent	Med	R-E.6	SBX-304	3
No MCP / external tool servers	High	R-F.1	ECO-301	3
No plugin authoring SDK/scaffold	High	R-F.2	ECO-302	3
No container/microVM plugin loaders	Med	R-F.3	ECO-303	3
Plugin signing UX friction	Med	R-F.4	ECO-304	3
Marketplace scanner static-only (no CVE)	Med	R-F.5	ECO-305	3
No graceful shutdown/drain	High	R-G.1	SRV-101	1
Async runs not durable	High	R-G.2	SRV-102	1
Webhook delivery in-process, no outbox	High	R-G.3	SRV-103	1
API keys no expiry/rotation/revocation	High	R-G.4	SRV-104	1
Rate limiter in-process only	High	R-G.5	SRV-201	2
Only 2/4 quota dims; no token quota	Med	R-G.6	SRV-202	2
Daily-cost window UTC-only	Med	R-G.7	SRV-203	2
FileApiKeyStore reads file per request	Med	R-G.8	SRV-302	3
No served OpenAPI; no WebSocket	Med	R-G.8	SRV-303	3
lib.rs ~86% inline test	Med	R-G.8	SRV-304	3
Idempotency full-file write per req	Med	R-G.8	SRV-305	3
Request-path unwraps	Low	R-G.8	SRV-306	3
In-process concurrency slots	Low	R-G.6	SRV-307	3
Postgres single client, no pool	High	R-H.1	WFL-101	1
Unbounded sub-workflow recursion	High	R-H.2	WFL-102	1
No TLS to Postgres	High	R-H.3	WFL-103	1
Event seq via MAX+1 race	Med	R-H.4	WFL-104	1
No loop/for-each	High	R-H.5	WFL-301	3
No dynamic fan-out	High	R-H.6	WFL-302	3
Checkpoint write-amp / no payload cap	Med	R-H.7	WFL-303	3
Event log no compaction / full load	Med	R-H.7	WFL-304	3
list() scans all checkpoints	Med	R-H.7	WFL-305	3
Timer/schedule poll O(N) / accuracy	Med	R-H.7	WFL-306	3
No activity progress events	Low	R-H.8	WFL-307	3
Event enum no versioning	Low	R-H.8	WFL-308	3
wovyr-runtime ai activity underspecified	Med	R-A.4	RUN-201	2
Sub-agent observability lost (NullSink, $0)	Low	R-B.1	RUN-202	2
No guardrails/moderation/PII	Med	R-I.1	SAF-201	2
No prompt template/versioning	Med	R-I.2	SAF-202	2
Default secrets plaintext	High	R-I.3	SEC-101	1
Audit query no time-range/pagination/scale	Med	R-I.4	SEC-301	3
Secret injection env-var leak surface	Low	R-I.5	SEC-302	3
Version 0.1.0 vs v0.3.0; no CHANGELOG	High	R-J.1	DX-101	1
No release automation / published image	High	R-J.2	DX-102	1
Dashboard + Windows absent from CI	High	R-J.3	DX-103	1
SDK parity (async py, retry, poll helper)	Med	R-J.4	DX-301	3
No coverage/benchmark tracking	Med	R-J.4	DX-302	3
SDK versioning strategy absent	Med	R-J.5	DX-303	3
commands.md out of sync	Med	R-J.6	DX-304	3
Docs status front-matter; quickstart; repo URL	Low	R-J.6	DX-305	3
Go/Java clients decision	Low	R-J.7	DX-306	3
Token in localStorage	High	R-K.1	UI-101	1
Zero UI test coverage	High	R-K.2	UI-102	1
No shared component abstraction	Med	R-K.3	UI-301	3
API types drift; string-built YAML; swallowed errors	Med	R-K.4	UI-302	3
No audit-log viewer	High	R-K.5	UI-303	3
Weak responsive/mobile	Med	R-K.6	UI-304	3
Accessibility thin	Med	R-K.7	UI-305	3
Playground / nav badges / i18n / icon sprite	Low	R-K.8	UI-306	3
No per-tenant metrics	Med	R-L.1	OBS-201	2
No queue-depth/DLQ gauges	Med	R-L.2	OBS-301	3
Uneven trace coverage	Low	R-L.3	OBS-302	3
No systemd/install; no upgrade runbook	High	R-L.4	DEP-301	3
No SLO burn; Helm HA/TLS; Terraform	Low	R-L.5	DEP-302	3

---

13. Revision History

Version	Date	Description
1.0.1	2026-07-15	Status updated Draft → In delivery: Phases 1–2 complete, Phase 3 partial (22/39), remainder re-prioritized through PRD-005; no content changed
1.0.0	2026-07-09	Initial AI-platform-maturity PRD: a 2026-07-09 five-front engineering audit’s ~90 findings mapped to 12 workstreams / testable requirements, phased into three ticket docs (RM-AIM-P1/P2/P3), with the distributed Scale-Out fold-in and a full findings→requirements→ticket traceability matrix