sparda-mcp 0.58.0 → 0.63.0

package/README.md

@@ -35,42 +35,47 @@ npx sparda-mcp badge        # a README badge: proven · coverage% · routes
 Under the hood it compiles your backend into one language-agnostic graph — the **Unified Behavior Graph (UBG)**, serialized as `.sparda/ubg.json` under the **SBIR** specification ([SPARDA Behavior IR](docs/SBIR_SPEC_V1.1.md)) — and every command is a pass over that graph.
 
 > [!IMPORTANT]  
-> **The Route-Compilation Proof — reproduce it yourself.** SPARDA compiles real open-source monsters to their behavior graph with **zero crashes**, each in **≈1–2 seconds**: Next.js *Dub* (579 routes), NestJS *Immich* (281), *MedusaJS* (477). It natively resolves deep Dependency Injection, external controllers, and Next.js handlers. One command clones them and re-measures on your machine:
+> **The Route-Compilation Proof — reproduce it yourself.** SPARDA compiles real open-source monsters to their behavior graph with **zero crashes**, each in **≈1–2 seconds**: Next.js _Dub_ (579 routes), NestJS _Immich_ (281), _MedusaJS_ (477). It natively resolves deep Dependency Injection, external controllers, and Next.js handlers. One command clones them and re-measures on your machine:
+>
 > ```bash
 > node bench/repro.mjs        # → bench/route-proof.json
 > ```
-> Honesty first: *compiling* a route is a parser result (the number above); *proving* it safe is a separate per-repo verdict — and most real apps come back **NOT_PROVEN**, which is the true state, not a failure. (Our full 25-repo corpus stress compiles **3,565 routes** at ~150 routes/s; that one needs the corpus checked out.)
+>
+> Honesty first: _compiling_ a route is a parser result (the number above); _proving_ it safe is a separate per-repo verdict — and most real apps come back **NOT_PROVEN**, which is the true state, not a failure. (Our full 25-repo corpus stress compiles **3,565 routes** at ~150 routes/s; that one needs the corpus checked out.)
 
 **What the graph unlocks — 100% local, deterministic, zero runtime dependencies, zero API key:**
 
-| Command | What it does |
-|---|---|
-| **`prove`** | *The whole trust verdict in one gesture* — proof + coverage + a shareable seal (`--json` / `--markdown`) |
-| **`apocalypse`** | *Prove the deploy* — no guard, invariant, transaction or aggregate boundary can be broken (SARIF + CI gate) |
-| **`badge`** | *The shareable artifact* — a self-contained SVG badge + README snippet (verdict · coverage · routes) |
-| **`dossier`** | *The public report* — one self-contained HTML page: verdict, risks, and SPARDA's own blind spots |
-| **`ubg`** | Compile the codebase to its behavior graph (Express · FastAPI · Next.js · NestJS · Medusa natively; **any** stack via OpenAPI) |
-| **`timeless`** | *Time-travel* — record a production request, replay it byte-identically, export the bug as a test |
-| **`heal`** | *Self-heal, proven* — bug → fix → the machine proves the fix is correct and breaks nothing |
-| **`mirror`** | *Execute the graph* — serve the compiled behavior over HTTP with no framework and no source |
-| **`init` / `dev`** | *Runtime, optional* — expose the graph to AI clients as a live MCP server (+ Twin, Immune, Evolution) |
+| Command            | What it does                                                                                                                                                                                                                         |
+| ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| **`prove`**        | _The whole trust verdict in one gesture_ — proof + coverage + a shareable seal (`--json` / `--markdown`)                                                                                                                             |
+| **`apocalypse`**   | _Prove the deploy_ — no guard, invariant, transaction or aggregate boundary can be broken (SARIF + CI gate)                                                                                                                          |
+| **`heal`**         | _Self-heal, **proven**_ — the gate Copilot Autofix doesn't have: a fix ships **only if** replay matches, `verify` still passes, and `apocalypse` finds no new risk / no dropped guard. Whoever wrote the fix, the machine judges it. |
+| **`badge`**        | _The shareable artifact_ — a self-contained SVG badge + README snippet (verdict · coverage · routes)                                                                                                                                 |
+| **`dossier`**      | _The public report_ — one self-contained HTML page: verdict, risks, and SPARDA's own blind spots                                                                                                                                     |
+| **`ubg`**          | Compile the codebase to its behavior graph (Express · FastAPI · Next.js · NestJS · Medusa natively; **any** stack via OpenAPI)                                                                                                       |
+| **`timeless`**     | _Time-travel_ — record a production request, replay it byte-identically, export the bug as a test                                                                                                                                    |
+| **`mirror`**       | _Execute the graph_ — serve the compiled behavior over HTTP with no framework and no source                                                                                                                                          |
+| **`init` / `dev`** | _Runtime, optional_ — expose the graph to AI clients as a live MCP server (+ Twin, Immune, Evolution)                                                                                                                                |
 
-The prover is the product. The MCP server is one *output* of the graph, not the point — SPARDA compiles the whole system's behavior, then proves, replays, heals, and (optionally) serves it.
+The prover is the product. The MCP server is one _output_ of the graph, not the point — SPARDA compiles the whole system's behavior, then proves, replays, heals, and (optionally) serves it.
 
-**Nomenclature:** **SBIR** is the specification (the format, like "JSON"); **UBG** is the compiled graph itself (the artifact, `ubg.json`). The MCP server is one *output* of the graph, not the product.
+**Nomenclature:** **SBIR** is the specification (the format, like "JSON"); **UBG** is the compiled graph itself (the artifact, `ubg.json`). The MCP server is one _output_ of the graph, not the product.
 
 ## Optional: expose the graph to AI clients (MCP runtime)
 
 Beyond proving, SPARDA can turn your running app into a live MCP server — the graph, executable, with write-safety and an immune layer. This is optional and separate from the prover above.
 
 1. **Scan + inject** — run once, from your app's directory:
+
    ```bash
    npx sparda-mcp init
    ```
+
    SPARDA parses your routes (AST), generates a marked `/mcp` router, injects it into
    your app (with a backup), and writes `sparda.json`. Every step is reversible.
 
 2. **Start your app, then start the bridge:**
+
    ```bash
    npx sparda-mcp dev
    ```
@@ -94,25 +99,31 @@ Beyond proving, SPARDA can turn your running app into a live MCP server — the
 ## Try the Standalone Demo
 
 To see SPARDA in action instantly without modifying your codebase:
+
 ```bash
 npx sparda-mcp demo
 ```
+
 This runs the entire MCP lifecycle (detect → parse → generate → inject → remove) on a bundled demo app in a temporary folder, in about 10 seconds. For the compiler itself, run `npx sparda-mcp ubg` then `apocalypse` on any Express/FastAPI app.
 
 ## Black Box Report
 
 SPARDA is designed as a local organism. To see what it remembers and how much compute it has recycled:
+
 ```bash
 npx sparda-mcp report
 ```
+
 This prints a terminal dashboard aggregating your exposed tools, write opt-ins, proof journal decisions, and crystallized composite tools.
 
 To write a self-contained, offline HTML dashboard at `.sparda/report.html`, append the `--html` flag:
+
 ```bash
 npx sparda-mcp report --html
 ```
 
 To output raw JSON for integration:
+
 ```bash
 npx sparda-mcp report --json
 ```
@@ -120,28 +131,35 @@ npx sparda-mcp report --json
 ## Deployment Proof: Apocalypse
 
 SPARDA's Behavior Graph is a formal model of your system. Instead of waiting for runtime failures or relying on static analysis vibes, you can statically prove the safety of your backend before any deployment:
+
 ```bash
 npx sparda-mcp apocalypse
 ```
+
 This command reads the compiled `.sparda/ubg.json` (with zero source code parsing at runtime) and discharges five static correctness obligations:
-* **Unguarded Mutation (Critical)**: Flags any mutation path that does not cross a security `guard`.
-* **Non-Atomic Aggregate Write (High)**: Flags when an API writes to multiple tables of the same Consistency Domain (Aggregate) outside a single transaction scope.
-* **Unvalidated Constrained Write (Medium)**: Flags writes into columns with declared invariants (CHECK, NOT NULL, UNIQUE — parsed from your `.sql` DDL **or `schema.prisma`**, Prisma enums included) without prior validation (Zod/Pydantic).
-* **Irreversible Observable Effect (High)**: Flags out-of-process actions (like Stripe charges) that happen alongside state writes without a structural compensation path (like a catch-refund).
-* **Aggregate Member Bypass (Info)**: Flags mutating a member table directly without routing through the aggregate root.
+
+- **Unguarded Mutation (Critical)**: Flags any mutation path that does not cross a security `guard`.
+- **Non-Atomic Aggregate Write (High)**: Flags when an API writes to multiple tables of the same Consistency Domain (Aggregate) outside a single transaction scope.
+- **Unvalidated Constrained Write (Medium)**: Flags writes into columns with declared invariants (CHECK, NOT NULL, UNIQUE — parsed from your `.sql` DDL **or `schema.prisma`**, Prisma enums included) without prior validation (Zod/Pydantic).
+- **Irreversible Observable Effect (High)**: Flags out-of-process actions (like Stripe charges) that happen alongside state writes without a structural compensation path (like a catch-refund).
+- **Aggregate Member Bypass (Info)**: Flags mutating a member table directly without routing through the aggregate root.
 
 To save your current graph as a safe baseline:
+
 ```bash
 npx sparda-mcp apocalypse --save-baseline
 ```
+
 Subsequent runs will diff the candidate graph against this baseline to detect regression vectors:
-* Deletion of any security `guard` (Critical).
-* Deletion of a database SQL invariant (High).
-* API blast radius expansion (Medium).
+
+- Deletion of any security `guard` (Critical).
+- Deletion of a database SQL invariant (High).
+- API blast radius expansion (Medium).
 
 If any Critical or High finding is found, `apocalypse` exits with a non-zero code to block your CI pipeline.
 
 **One step in your workflow — findings land in the GitHub Security tab (SARIF):**
+
 ```yaml
 - uses: zyx77550/sparda@main
   with:
@@ -159,11 +177,13 @@ npx sparda-mcp timeless export <id>    # the production bug is now a vitest test
 ```
 
 Recording is two lines in your app (ESM), with deterministic sampling and GDPR redaction built in:
+
 ```js
 import { getFlightBox } from 'sparda-mcp/src/flight/box.js';
-const box = getFlightBox(); box.arm();
-app.use(box.middleware({ sample: 100 }));   // 1 request in 100; passwords/tokens redacted by default
-const db = box.wrapClient(pgPool);           // your query client, tapped
+const box = getFlightBox();
+box.arm();
+app.use(box.middleware({ sample: 100 })); // 1 request in 100; passwords/tokens redacted by default
+const db = box.wrapClient(pgPool); // your query client, tapped
 ```
 
 The closed loop nobody else has: **production bug → recorded flight → failing test → AI writes the fix → `apocalypse` proves the fix breaks no guard, invariant or transaction → deploy.** Replay is per-request (concurrent-race capture is out of scope for v1 — stated, not hidden).
@@ -180,7 +200,7 @@ npx sparda-mcp heal <flightId> --check --expect '{"status":404}'
 
 The brief is built from the graph itself — it hands the fixer the handler's `file:line`, the capabilities the fix must not grow, and the guards it must not remove. Then the **gate** — the actual product — proves the fix on three axes at once:
 
-1. **Behavior** — lenient replay of the recorded flight (same deterministic inputs) now produces the *expected* response, not the recorded bug. The fix may reformulate a query (the tap is relabeled, allowed); it may **not** change the effect order or kinds.
+1. **Behavior** — lenient replay of the recorded flight (same deterministic inputs) now produces the _expected_ response, not the recorded bug. The fix may reformulate a query (the tap is relabeled, allowed); it may **not** change the effect order or kinds.
 2. **Compiler laws** — `verify` still passes: the graph is still sound and deterministic.
 3. **No regression** — `apocalypse` diff against the frozen pre-fix graph: zero new critical/high findings, no guard removed, no blast radius grown.
 
@@ -188,7 +208,7 @@ The brief is built from the graph itself — it hands the fixer the handler's `f
 ✓ HEALED & PROVEN — same recorded inputs, correct output, zero law broken, zero protection lost. Ship it.
 ```
 
-The gate is honest in both directions: an unfixed bug, or a "fix" that silently drops a guard, keeps it **closed** (exit 1). This is the difference between an AI that writes plausible code and a system that *proves* the code is correct — the trust layer the agent era is missing.
+The gate is honest in both directions: an unfixed bug, or a "fix" that silently drops a guard, keeps it **closed** (exit 1). This is the difference between an AI that writes plausible code and a system that _proves_ the code is correct — the trust layer the agent era is missing.
 
 ## Any Backend On Earth: OpenAPI Lowering
 
@@ -233,7 +253,7 @@ To undo everything: **`npx sparda-mcp remove`** restores your code byte-for-byte
 5. **Nothing leaves your machine.** No telemetry to us, no cloud, local key auth, 4 exact-pinned dependencies.
 6. **What it learns is never lost.** Diagnoses, descriptions, settings — versioned with your git, surviving every re-init.
 
-What we *don't* promise: the honest limits in [docs/SECURITY.md](./docs/SECURITY.md).
+What we _don't_ promise: the honest limits in [docs/SECURITY.md](./docs/SECURITY.md).
 
 ## How it works
 
@@ -250,34 +270,52 @@ What we *don't* promise: the honest limits in [docs/SECURITY.md](./docs/SECURITY
 ## What SPARDA gives your AI
 
 ### Operate, not just read
+
 Every route becomes a tool that runs against your live process — real auth, real data,
 warm connections. One call to **`sparda_get_context`** hands the AI the whole living
 picture: enabled tools, suggested workflows, runtime telemetry, quarantine state, and
 immune memory — so every session resumes where the last one stopped.
 
+### Prove the edit before you commit — the one check an LLM can't do to itself
+
+The AI just edited a route. Did it quietly drop a guard? It calls **`sparda_prove`** and
+finds out **now**, not in a CI run later. The tool recompiles the app to its behavior graph,
+discharges the same static obligations as `sparda apocalypse`, and returns a deterministic
+verdict — the exact word the CLI and badge emit, so it can never over-claim (a low-coverage
+clean app reads `SURFACE`, never a bare `PROVEN`). Save a baseline once
+(`sparda apocalypse --save-baseline`) and every later `sparda_prove` flags any finding with
+`regression: true` — the guard your edit removed, the route it dropped, the blast radius it
+grew. That's _"AI writes. SPARDA proves."_ inside the edit loop. Clients that list MCP prompts
+also get the **`prove-my-edit`** workflow.
+
 ### Write-safety: the AI can't write until you say so
+
 - Writes (POST/PUT/DELETE) ship **disabled**. Enable them per tool in `sparda.json`; your choice survives every re-init.
 - An enabled write is **never executed on the first call**. SPARDA returns an `awaiting_confirmation` envelope — a single-use token plus a preview of the action — and commits only after an explicit confirm step.
 - When your client supports MCP elicitation, that confirmation prompt appears **in the AI's own UI**.
 - **Proof-after-write**: every successful write is followed by a read-back of the same resource, so the AI — and you — see the real effect, not a hopeful guess.
 
 ### Your app defends itself — zero LLM on the hot path
+
 - **Quarantine.** A tool that returns 3 consecutive 5xx is quarantined: further calls get a `503` with a reason and a retry delay instead of hammering your broken route. After a cooldown it half-opens for a single probe.
 - **Latency & anomaly flags.** The router learns each route's baseline and flags deviations locally, in a few lines of math.
 - **Adaptive diagnosis, only on surprise.** A genuinely new failure wakes your AI client's own model to diagnose it once; the diagnosis is cached as an "antibody" in `sparda.json`, so the same failure later costs zero tokens. Cloning your code doesn't clone its immune memory.
 
 ### A free intelligence layer, zero API key
+
 On first connection your AI client's own model (via MCP sampling) rewrites raw routes
 into business-language tool descriptions and proposes multi-step workflows — cached in
 `sparda.json` and exposed as MCP prompts. Nothing to configure, nothing to pay.
 
 ### It gets cheaper the more you use it
+
 - **Response recycling.** When a read keeps returning the same answer, SPARDA serves the next identical call straight from memory — without touching your host app. Reads only; writes always hit the host.
 - **A recycling gauge.** `GET /mcp/stats` counts how many calls were answered from SPARDA's own knowledge vs. how many paid the host route. It reads 0% on day one and fills with usage — a measure, never a promise.
 
 ### Tools nobody wrote — Labs, opt-in, default OFF
+
 Turn it on with `"labs": { "recordSequences": true }` in `sparda.json`. SPARDA then
-notices when one tool's output feeds the next tool's input and records the *circuit* —
+notices when one tool's output feeds the next tool's input and records the _circuit_ —
 structure only (tool names, argument names, counts), never your data. A read-only
 circuit seen enough times **crystallizes into a composite tool**, announced
 mid-session: one call runs the whole chain, auto-feeding each step from the previous
@@ -285,11 +323,13 @@ step's real response. Write routes are never absorbed — their per-call confirm
 always stands.
 
 ### Living context & telemetry
+
 `GET /mcp/stats` (per-tool calls/errors, tool "purity", quarantine state) and
 `GET /mcp/events` (errors, latency anomalies, cached diagnoses) expose exactly what
 your app is doing — surfaced to the AI as live notifications.
 
 ## Built for AI clients: the bundled Skill
+
 SPARDA ships with an Agent Skill ([`SKILL.md`](./SKILL.md)) that teaches any compatible
 AI client how to drive a SPARDA server to its **full potential** — call
 `sparda_get_context` first, exploit response recycling, honor quarantine, prefer
@@ -307,6 +347,7 @@ runtime, so the guidance never goes stale.
 - **FastAPI** (Python >= 3.9) — AST-based router injection.
 
 ## Security posture (honest)
+
 - 4 runtime dependencies, exact-pinned.
 - **Dynamic Local Key Resolution.** The generated router contains no baked secrets. It resolves authorization keys at runtime from the `SPARDA_LOCAL_KEY` environment variable or the local gitignored `.sparda/key` file, and fails closed (503) when neither is found. For custom production or staging setups, you can override this behavior by exposing `SPARDA_LOCAL_KEY` in your environment.
 - Local key on every router call; self-reference loop protection; 30s timeouts; 8 KB output truncation.
@@ -316,18 +357,21 @@ runtime, so the guidance never goes stale.
 Full threat model and known gaps: [docs/SECURITY.md](./docs/SECURITY.md).
 
 ## Documentation
+
 - [docs/ARCHITECTURE.md](./docs/ARCHITECTURE.md) — how `init`, the injected router, and the bridge fit together, plus the `sparda.json` schema.
 - [docs/SECURITY.md](./docs/SECURITY.md) — threat model, defenses, and honest known gaps.
 - [docs/TESTING.md](./docs/TESTING.md) — how the promises above are kept honest in CI.
 - [docs/ERRORS.md](./docs/ERRORS.md) — the error knowledge base.
 
 ## Beyond the open core
+
 SPARDA is free, including in production (see License). Team-scale capabilities —
 fine-grained per-person access policies and a signed, tamper-evident audit log — are
 planned for a future paid tier. The open core stands on its own; nothing here is
 crippled to upsell you.
 
 ## License
+
 [Business Source License 1.1](./LICENSE) — free to use, including in production.
 You may not resell SPARDA or offer it as a competing commercial service.
 Each version converts to Apache 2.0 four years after its release.

package/SKILL.md

@@ -13,11 +13,11 @@ description: >-
 
 A SPARDA server is driven by a compiled **Unified Behavior Graph (UBG)** — the graph SPARDA's compiler produces from the host application's states, transitions, permissions, and side-effects, serialized under the **SBIR** specification. Instead of exposing raw, disconnected endpoints, SPARDA compiles the app into a deterministic behavioral model. The local **SPARDA Runtime** dynamically executes this graph inside the live host process, powering the MCP interface, the Twin simulation clone, and the Immune system offline.
 
-> This skill covers the **runtime** (driving a live MCP server) — the live half of SPARDA's trust layer: *"AI writes. SPARDA proves."* The same graph also powers dev-time proof commands you run in the app's repo — `sparda review` (the behavior diff of a PR), `apocalypse` (prove the deploy), `timeless` (record/replay a request), `heal` (prove a fix), `mirror` (serve the graph), `ubg` (compile), `verify` (prove the compiler's laws). Those are CLI, not MCP tools; see the project README.
+> This skill covers the **runtime** (driving a live MCP server) — the live half of SPARDA's trust layer: _"AI writes. SPARDA proves."_ The same graph also powers dev-time proof commands you run in the app's repo — `sparda review` (the behavior diff of a PR), `apocalypse` (prove the deploy), `timeless` (record/replay a request), `heal` (prove a fix), `mirror` (serve the graph), `ubg` (compile), `verify` (prove the compiler's laws). Those are CLI, not MCP tools; see the project README.
 
 ## Rule 0 — call `sparda_get_context` first, every session
 
-Before anything else, call **`sparda_get_context`** (no params). It returns the *live* state of the SPARDA Behavior Graph:
+Before anything else, call **`sparda_get_context`** (no params). It returns the _live_ state of the SPARDA Behavior Graph:
 
 - the active routes/tools, workflows, and type-propagated schemas;
 - `runtime` — current stats (calls, errors, quarantine states, Twin mode active);
@@ -35,19 +35,45 @@ Read it to orient yourself inside the graph. `sparda_info` gives a lighter summa
   parameter schema was only partially inferred — pass arguments carefully.
 - **Meta-tools** — `sparda_get_context`, `sparda_info`,
   `sparda_list_disabled_tools`, `sparda_confirm`.
+- **The proof tool** — `sparda_prove`. Call it **after you edit a route, before you
+  commit** — see _Prove your own edit_ below. It's the one check you can't do to
+  yourself by re-reading your code.
 - **Composite tools** — labelled `[Labs circuit ×N]`, `readOnly`. One call runs a
-  whole proven multi-step chain (see *Crystallized circuits* below).
+  whole proven multi-step chain (see _Crystallized circuits_ below).
 
 Only **enabled** tools appear. Write tools are hidden until the user opts in, so a
-missing write is a config state, not an error — see *Writing safely*.
+missing write is a config state, not an error — see _Writing safely_.
+
+## Prove your own edit — call `sparda_prove` before you commit
+
+This is the tool an LLM needs most and can least fake. After you edit a route,
+**call `sparda_prove`** — it recompiles the app to its behavior graph and discharges
+the same static obligations as `sparda apocalypse` (unguarded mutation, non-atomic
+aggregate write, unvalidated constrained write), then returns a deterministic verdict.
+
+- **Focus it.** Pass `route` with the method+path you just touched (e.g.
+  `{ "route": "DELETE /orders" }`) to narrow the finding list. The **verdict still
+  reflects the whole app** — a filter never buys you a greener light.
+- **Read the verdict honestly.** `PROVEN` / `PARTIAL` are safe to commit. `SURFACE`
+  and `NO_PROOF` mean SPARDA _couldn't resolve enough to prove it_ — that is
+  "unknown", **never** a pass. The word is the exact one the CLI and badge emit; it
+  physically cannot over-claim.
+- **The regression check is the point.** If a baseline was saved
+  (`sparda apocalypse --save-baseline` on a known-good state), any finding with
+  `regression: true` means _your edit_ removed a guard, dropped a route, or grew the
+  blast radius vs the last proven state. Fix those before committing — this is the
+  check you cannot perform by re-reading your own diff.
+- Clients that list MCP prompts also see **`prove-my-edit`**, the built-in workflow
+  that walks these steps.
 
 ## Exploit the intelligence layer (this is the "full potential")
 
 **1. Response-recycling flywheel — make repeated reads free.**
-When the *same* read tool returns a byte-identical result for the *same* arguments
+When the _same_ read tool returns a byte-identical result for the _same_ arguments
 **3 times within 30 seconds**, SPARDA serves the next identical call straight from
 RAM (`servedByFlywheel: true`) **without touching the host app**. So:
-- Don't fear repeating stable GETs — repetition is what *activates* the cache.
+
+- Don't fear repeating stable GETs — repetition is what _activates_ the cache.
 - Don't bolt your own client-side cache on top; you'd hide the signal that lets
   SPARDA recycle, and you'd lose freshness control.
 - Watch `recycling.flywheel.servedFromMemory` climb in context — that's free work.
@@ -55,7 +81,7 @@ RAM (`servedByFlywheel: true`) **without touching the host app**. So:
 
 **2. Circuit-breaker / quarantine — stop hammering a sick backend.**
 After **3 consecutive 5xx** on a tool, SPARDA quarantines it: subsequent calls
-return **HTTP 503** with `reason` and `retryInMs` *instead of* hitting the failing
+return **HTTP 503** with `reason` and `retryInMs` _instead of_ hitting the failing
 host. Honor `retryInMs` — do not retry-loop. Check `runtime.quarantine` in context
 before depending on a tool. After a cooldown (~60s) the tool half-opens for one
 probe; one more 5xx re-quarantines it.
@@ -68,7 +94,7 @@ single **composite tool** for that chain and announces it mid-session via
 call and it's marked read-only. (Writes are never absorbed into a circuit.)
 
 **4. Adaptive immunity — read the diagnosis before retrying.**
-Repeated, unfamiliar failures trigger a *one-shot* LLM diagnosis that SPARDA caches
+Repeated, unfamiliar failures trigger a _one-shot_ LLM diagnosis that SPARDA caches
 as an "antibody" (keyed by `source|tool|status`). Recurrences reuse the cached
 diagnosis at zero cost. When an error event carries a diagnosis, **read it** and
 adapt — don't blindly retry the same call.
@@ -78,11 +104,13 @@ an `immune` event in `/mcp/events`. Treat it as a hint to back off or warn the u
 
 **6. Twin Simulation Mode — practice safely on a clone.**
 When `/mcp/stats` or `sparda_get_context.runtime` contains `"twin": true`, you are connected to a safe, in-memory mock clone of the application.
+
 - All GET reads return learned exemplars (observed response shapes and mock values).
 - All write tools return simulated `202` echoes but do not write to database or external APIs.
 - Use this twin mode to practice multi-step workflows, debug tool sequences, and test your plans without touching the live production backend.
 
 **7. Grammar & Evolution — discover optimal workflows.**
+
 - You can query or contribute to the app's grammar (`.sparda/grammar.json`). The grammar maps valid sequences of tool calls (edges).
 - Running `sparda evolve` mutates and runs candidate chains against the twin. The successful evolved sequences are suggested as mid-session workflows.
 
@@ -151,7 +179,8 @@ Writes are **disabled by default**. The protocol is not optional:
 - **Learn exemplars** → Start your live app and run `sparda twin --learn` to fetch actual response data and construct `.sparda/twin.json` locally.
 
 ---
-*This skill ships with `sparda-mcp` and is regenerated from SPARDA's capability
+
+_This skill ships with `sparda-mcp` and is regenerated from SPARDA's capability
 surface each release, so it tracks new tools and behaviors. The **live, per-project**
 tool list, stats, and workflows always come from `sparda_get_context` at runtime —
-trust it over any static list.*
+trust it over any static list._

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "sparda-mcp",
-  "version": "0.58.0",
+  "version": "0.63.0",
   "mcpName": "io.github.zyx77550/sparda-mcp",
   "description": "Compile backends to behavior graphs. Statically prove security, simulate APIs, and replay bugs.",
   "type": "module",
@@ -26,7 +26,9 @@
     "lint": "eslint .",
     "lint:fix": "eslint . --fix",
     "format": "prettier --write \"**/*.{js,cjs,mjs}\"",
-    "format:check": "prettier --check \"**/*.{js,cjs,mjs}\""
+    "format:check": "prettier --check \"**/*.{js,cjs,mjs}\"",
+    "bench:check": "node bench/check-readme.mjs",
+    "mutation": "node tests/mutation/run.mjs"
   },
   "files": [
     "src",
@@ -55,8 +57,7 @@
     "@babel/parser": "7.26.5",
     "@babel/traverse": "7.26.5",
     "@clack/prompts": "0.9.1",
-    "@modelcontextprotocol/sdk": "1.29.0",
-    "js-tokens": "^10.0.0"
+    "@modelcontextprotocol/sdk": "1.29.0"
   },
   "devDependencies": {
     "@eslint/js": "^9.39.4",

package/src/commands/stitch.js

@@ -0,0 +1,58 @@
+// commands/stitch.js — cross-repo / cross-service behavior stitching. Point it at two or more app
+// directories (microservices of the same system); it compiles each, joins the outbound HTTP calls
+// of one to the entrypoints of another, and reports the trust boundaries + any cross-service BOLA
+// the join reveals — a finding no mono-repo tool can produce. Quorum sensing: each service's graph
+// is a signal; the collective behavior emerges from reading them together.
+import path from 'node:path';
+import { compileUBG } from '../ubg/compile.js';
+import { canonicalizeGraph } from '../ubg/schema.js';
+import { checkGraph } from '../ubg/apocalypse.js';
+import { stitchServices } from '../ubg/stitch.js';
+
+export async function runStitch(opts, dirs) {
+  if (!dirs || dirs.length < 2) {
+    console.error(
+      'Usage: sparda stitch <dir1> <dir2> [...]  — two or more service/app directories',
+    );
+    process.exitCode = 1;
+    return;
+  }
+  const services = [];
+  for (const d of dirs) {
+    const abs = path.resolve(opts.cwd, d);
+    try {
+      const { graph } = compileUBG(abs, { write: false });
+      const g = canonicalizeGraph(graph);
+      const { findings } = checkGraph(g);
+      services.push({ name: path.basename(abs), graph: g, findings });
+    } catch (e) {
+      console.error(`  ✗ ${d}: ${e.message.slice(0, 70)}`);
+    }
+  }
+
+  const { edges, findings } = stitchServices(services);
+
+  if (opts.json) {
+    console.log(
+      JSON.stringify({ services: services.map((s) => s.name), edges, findings }, null, 2),
+    );
+    return;
+  }
+
+  console.log(`\nSPARDA · stitch — ${services.map((s) => s.name).join(' · ')}`);
+  console.log('─'.repeat(52));
+  if (!edges.length) {
+    console.log(
+      '  no cross-service calls resolved (targets may be dynamic or unrelated)',
+    );
+  } else {
+    console.log(`  ${edges.length} cross-service call(s):`);
+    for (const e of edges)
+      console.log(`    ${e.fromService} → ${e.toService}  ${e.method} ${e.path}`);
+  }
+  if (findings.length) {
+    console.log(`\n  ◐ ${findings.length} cross-service advisory(ies):`);
+    for (const f of findings) console.log(`    ${f.message}`);
+  }
+  console.log('');
+}

package/src/index.js

@@ -40,6 +40,34 @@ const opts = {
   cwd: path.resolve(process.cwd(), getOpt('dir', null) ?? getOpt('cwd', null) ?? '.'),
 };
 
+// Per-command help — `sparda <cmd> --help` (or -h) prints usage and exits, instead of running
+// the command (the old behavior: `--help` was ignored and the command just ran). Every command
+// reads opts.cwd, so `--dir <path>` works on all of them; only per-command extras are listed.
+const HELP = {
+  prove: `sparda prove [--dir <path>] [--json | --markdown]\n  The whole trust verdict: proof + coverage + a shareable seal.\n  --markdown  emit a sticky-PR-comment body (used by the GitHub Action).`,
+  apocalypse: `sparda apocalypse [--dir <path>] [--sarif] [--save-baseline]\n  Prove the deploy — exit 1 on any critical/high finding.\n  --sarif         also write .sparda/apocalypse.sarif for the Security tab.\n  --save-baseline freeze this graph; later runs diff against it.`,
+  ubg: `sparda ubg [--dir <path>] [--json] [--out <file>] [--openapi <spec>]\n  Compile the codebase to its Unified Behavior Graph (.sparda/ubg.json).`,
+  stitch: `sparda stitch <dir1> <dir2> [...] [--json]\n  Cross-service proof: join one service's outbound HTTP calls to another's routes,\n  surface cross-service trust boundaries + BOLA no mono-repo tool can see.`,
+  badge: `sparda badge [--dir <path>] [--out <file>] [--json]\n  Emit a shareable SVG badge + README snippet (verdict · coverage · routes).`,
+  dossier: `sparda dossier [--dir <path>] [--json]\n  Render the whole proof as one self-contained, shareable HTML page.`,
+  review: `sparda review [--dir <path>] --base <ref> [--json | --markdown]\n  Semantic behavior diff of the working tree vs a base git ref.`,
+  init: `sparda init [--dir <path>] [--yes]\n  Scan the app, generate & inject the reversible /mcp router.`,
+  dev: `sparda dev [--dir <path>] [--port <n>]\n  Run the MCP stdio bridge (connect an AI client).`,
+  remove: `sparda remove [--dir <path>] [--yes]\n  Remove SPARDA from this project — byte-for-byte clean git diff.`,
+  heal: `sparda heal <flightId> [--check] [--expect <json>] [--agent <cli>]\n  Turn a production bug into a fix brief + a proof gate.`,
+  timeless: `sparda timeless [list | replay <id> | export <id>]\n  Record / replay a production request byte-identically.`,
+  mirror: `sparda mirror [--dir <path>] [--port <n>]\n  Serve the compiled graph over HTTP — no framework, no source.`,
+};
+if (flags.has('--help') || rest.includes('-h')) {
+  if (cmd && HELP[cmd]) console.log(HELP[cmd]);
+  else if (cmd)
+    console.log(
+      `No detailed help for \`${cmd}\`. Run \`sparda\` for the full command list.`,
+    );
+  else console.log('Run `sparda` for the command list, or `sparda <command> --help`.');
+  process.exit(0);
+}
+
 try {
   switch (cmd) {
     case 'demo': {
@@ -121,6 +149,14 @@ try {
       await runBadge(opts);
       break;
     }
+    case 'stitch': {
+      const { runStitch } = await import('./commands/stitch.js');
+      await runStitch(
+        opts,
+        rest.filter((a) => !a.startsWith('--')),
+      );
+      break;
+    }
     case 'apocalypse': {
       const { runApocalypse } = await import('./commands/apocalypse.js');
       await runApocalypse(opts);
@@ -201,6 +237,7 @@ PROVE — the point
   review       Semantic diff of a PR vs a base ref (--base main / --json / --markdown)
   blindspots   Map SPARDA's own blindness — every unseen route/effect/guard, ranked (--json)
   badge        Emit a shareable SVG badge + README snippet (verdict · coverage · routes)
+  stitch       Cross-service proof — join N repos' graphs, find boundary-crossing BOLA
   dossier      Render the whole proof as one self-contained HTML page anyone can read
   verify       Prove the compiler's own laws (determinism, soundness, round-trip)
   heal         Turn a production bug into a fix brief + a gate (--check / --expect / --agent)

package/src/server/stdio.js

@@ -26,8 +26,39 @@ import { createSpardaEngine } from './engine.js';
 import { initiateWrite, preapproveWrite, confirmWrite } from './confirmation.js';
 import { resolveContext, injectContext, fingerprintContext } from './context-carrier.js';
 import { resolveSpardaKey } from '../generator/manifest.js';
+import { compileUBG } from '../ubg/compile.js';
+import { canonicalizeGraph } from '../ubg/schema.js';
+import { checkGraph, diffGraphs, verdictOf, verdictState } from '../ubg/apocalypse.js';
+import { surveyBlindspots } from '../ubg/blindspots.js';
 
 const EVENT_POLL_MS = Number(process.env.SPARDA_EVENT_POLL_MS ?? 5000);
+
+// Built-in MCP prompts (workflows) served by every SPARDA server, regardless of the app's own
+// inferred workflows. `prove-my-edit` is the discoverability surface for `sparda_prove`: the
+// workflow an editing agent lists and follows to check its OWN change before committing — the
+// one check an LLM cannot do to itself by re-reading its code.
+export const BUILTIN_WORKFLOWS = [
+  {
+    name: 'prove-my-edit',
+    description:
+      "Prove an edit didn't break a guard before you commit — the check an LLM can't do to itself by re-reading its own code.",
+    steps: [
+      'Call sparda_prove. If you just edited one route, pass route with its method+path (e.g. "DELETE /orders") to focus the finding list — the verdict still reflects the whole app.',
+      'Read the verdict: PROVEN / PARTIAL are safe to commit. SURFACE / NO_PROOF mean SPARDA could not resolve enough to prove it — treat that as "unknown", never as a pass.',
+      'Fix every finding with regression:true first — that edit removed a guard, dropped a route, or grew the blast radius vs the last proven baseline.',
+      'If baselined is false, run `sparda apocalypse --save-baseline` once on a known-good state so future sparda_prove calls can catch regressions.',
+    ],
+  },
+];
+
+// App-inferred workflows (carry-over, sacred) win on a name clash; built-ins fill the rest.
+export function mergeWorkflows(appWorkflows) {
+  const out = [...(appWorkflows ?? [])];
+  for (const b of BUILTIN_WORKFLOWS) {
+    if (!out.some((w) => w.name === b.name)) out.push(b);
+  }
+  return out;
+}
 // Advertised MCP server version — read from package.json so it can never drift from the
 // published package (it was pinned at a stale '0.5.2' for many releases).
 const SPARDA_VERSION = (() => {
@@ -340,6 +371,21 @@ export async function startStdioBridge({ cwd, portOverride }) {
           'Call this FIRST. Returns the full living context of this app: every tool with its description, known workflows, runtime telemetry (per-tool calls/errors/latency), quarantined tools, and the immune memory of past diagnosed failures. Lets any AI session resume exactly where the previous one stopped.',
         inputSchema: { type: 'object', properties: {} },
       },
+      {
+        name: 'sparda_prove',
+        description:
+          'Prove this app is safe to deploy — NOW, before you commit. Compiles the current source to its behavior graph and discharges the static proof obligations (unguarded mutation, non-atomic aggregate write, unvalidated constrained write). If a baseline was saved (`sparda apocalypse --save-baseline`), it ALSO diffs against it: any finding flagged `regression:true` means your edit removed a guard, dropped a route, or grew the blast radius vs the last proven state — the check an LLM cannot do to itself by re-reading code. Returns a deterministic verdict (PROVEN / PARTIAL / SURFACE / NO_PROOF / RISKY / NOT_PROVEN), the coverage % it resolved, and every finding with its route. The verdict is the exact word `sparda apocalypse` and the badge emit; it never over-claims (a low-coverage clean app reads SURFACE/PARTIAL, never a bare PROVEN). Pass `route` (e.g. "DELETE /orders") to focus the finding list on the route you just edited — the verdict still reflects the whole app.',
+        inputSchema: {
+          type: 'object',
+          properties: {
+            route: {
+              type: 'string',
+              description:
+                'Optional. Substring of a route label (method + path, e.g. "POST /invoices") to filter the findings to the route you just edited. Omit to see the whole app.',
+            },
+          },
+        },
+      },
       {
         name: 'sparda_confirm',
         description:
@@ -360,14 +406,14 @@ export async function startStdioBridge({ cwd, portOverride }) {
   }));
 
   server.setRequestHandler(ListPromptsRequestSchema, async () => ({
-    prompts: (manifest.semantic?.workflows ?? []).map((w) => ({
+    prompts: mergeWorkflows(manifest.semantic?.workflows).map((w) => ({
       name: w.name,
       description: w.description,
     })),
   }));
 
   server.setRequestHandler(GetPromptRequestSchema, async (req) => {
-    const w = (manifest.semantic?.workflows ?? []).find(
+    const w = mergeWorkflows(manifest.semantic?.workflows).find(
       (x) => x.name === req.params.name,
     );
     if (!w) throw new Error(`unknown prompt: ${req.params.name}`);
@@ -411,6 +457,9 @@ export async function startStdioBridge({ cwd, portOverride }) {
         ),
       );
     }
+    if (name === 'sparda_prove') {
+      return text(JSON.stringify(proveApp(cwd, { route: args.route }), null, 2));
+    }
     if (name === 'sparda_get_context') {
       const headers = hostHeaders(key, ctx);
       const [stats, events] = await Promise.all([
@@ -1081,6 +1130,76 @@ function text(t) {
   return { content: [{ type: 'text', text: t }] };
 }
 
+// The proof layer, served live over MCP — same compile + obligations as `sparda apocalypse`,
+// so an AI can prove its own edit structurally the moment it writes it, not a CI run later.
+// Reuses `verdictState` verbatim: this tool physically cannot show a word the CLI/badge won't
+// (the no-false-PROVEN invariant is one function, shared). Read-only (write:false); an
+// AI-initiated call, never on the host's request path, so Law 1 holds.
+//   opts.route  — substring; keep only findings on routes whose label matches (focus the
+//                 answer on the route the AI just edited, e.g. "DELETE /orders").
+export function proveApp(cwd, { route } = {}) {
+  let canonical, report;
+  try {
+    const compiled = compileUBG(cwd, { write: false });
+    canonical = canonicalizeGraph(compiled.graph);
+    report = compiled.report;
+  } catch (err) {
+    // A parser gap is NOT a pass — say so honestly, never a silent green.
+    return {
+      verdict: 'NO_PROOF',
+      provable: false,
+      note: `SPARDA could not compile this app's surface (${err.message}). An uncompiled app proves nothing — this is NOT a pass.`,
+    };
+  }
+  const { findings: staticFindings, obligations } = checkGraph(canonical);
+
+  // Baseline diff — the regression check that IS apocalypse's edge over a plain linter: did
+  // this edit REMOVE a guard, drop a route, or grow the blast radius vs the last saved proof?
+  // Only if a baseline was recorded (`sparda apocalypse --save-baseline`); absent → static only.
+  let diffFindings = [];
+  let baselined = false;
+  const baselinePath = path.join(cwd, '.sparda', 'ubg.baseline.json');
+  if (fs.existsSync(baselinePath)) {
+    try {
+      const baseline = JSON.parse(fs.readFileSync(baselinePath, 'utf8'));
+      diffFindings = diffGraphs(baseline, canonical).findings;
+      baselined = true;
+    } catch {
+      // unreadable baseline → degrade to static-only, never crash the AI's call
+    }
+  }
+
+  const all = [...staticFindings, ...diffFindings];
+  const label = (id) => canonical.nodes.find((n) => n.id === id)?.label ?? id;
+  let findings = all.map((f) => ({
+    severity: f.severity,
+    rule: f.rule,
+    route: label(f.entrypoint),
+    regression: diffFindings.includes(f),
+    message: f.message,
+  }));
+  if (route) {
+    const needle = route.toLowerCase();
+    findings = findings.filter((f) => f.route.toLowerCase().includes(needle));
+  }
+
+  const blind = surveyBlindspots(canonical, report);
+  // The verdict word always reflects the WHOLE app (a route filter narrows the finding list,
+  // never the safety claim — an AI must never read "PROVEN" because it hid the rest).
+  const verdict = verdictOf(all, canonical, { coverage: blind.coverage.ratio });
+  return {
+    verdict: verdictState(verdict),
+    provable: verdict.provable,
+    coverage: Math.round(blind.coverage.ratio * 100) / 100,
+    baselined,
+    obligations,
+    counts: verdict.counts,
+    findings,
+    blindspots: { surface: blind.surface, coverage: blind.coverage.ratio },
+    ...(route ? { scopedTo: route } : {}),
+  };
+}
+
 // AI-facing message when the Signal-2 gate (Brief #1) refuses a `sparda_confirm`. Each reason
 // tells the model what to do next without implying it can approve the write itself.
 function signal2Denial(reason) {

package/src/ubg/apocalypse.js

@@ -68,6 +68,10 @@ export function checkGraph(graph) {
   const g = indexGraph(graph);
   const findings = [];
   let obligations = 0;
+  // Resource tables by name — the input to BolaRay ownership-model inference (O7 below).
+  const statesByTable = new Map();
+  for (const n of g.nodes.values())
+    if (n?.kind === 'state' && n.meta?.table) statesByTable.set(n.meta.table, n);
   // Behavior polarity (ADR-036): per entrypoint, a ternary vector over the same
   // obligations checked below — +1 protection present, 0 not applicable, -1
   // violated. Built HERE so a -1 is literally the same condition as the finding
@@ -208,9 +212,15 @@ export function checkGraph(graph) {
       );
       vec.aggregate = touchesRoot && vec.aggregate !== -1 ? 1 : -1;
       if (!touchesRoot) {
+        // ADVISORY (E-046): a direct member-table write is a design-smell, not a proven
+        // violation — many ORMs/apps legitimately write members directly. On a real
+        // schema-rich app it fires in bulk (dub: 174), so it points a human at the pattern,
+        // it never gates the verdict. (Surfaced once the split-schema state layer became
+        // visible; making it hard would have flooded every folder-schema app.)
         findings.push({
           rule: 'AGGREGATE_MEMBER_BYPASS',
           severity: 'info',
+          advisory: true,
           entrypoint: ep.id,
           message: `${ep.label} mutates member table "${state.meta.table}" of aggregate "${state.meta.consistencyDomain}" without going through its root`,
           evidence: [w.stateId],
@@ -238,12 +248,27 @@ export function checkGraph(graph) {
       const tables = [
         ...new Set(idScoped.map((n) => n.meta.table).filter(Boolean)),
       ].sort();
+      // BolaRay (CCS 2024) step 1: infer the ownership MODEL each accessed table SHOULD carry
+      // (from its declared columns/FKs), so the advisory names the scope the access is MISSING
+      // instead of a vague "verify authorization". Still advisory (soundness unchanged): the
+      // schema tells us the model, never the runtime intent — that gap is why O7 never gates.
+      const ownership = tables.map((t) => ({
+        table: t,
+        ...(inferOwnershipModel(t, statesByTable) ?? { model: null, key: null }),
+      }));
+      const hints = ownership
+        .filter((o) => o.model)
+        .map((o) => `${o.table} should be ${o.model} (${o.key})`);
       findings.push({
         rule: 'OBJECT_SCOPE_UNPROVEN',
         severity: 'info',
         advisory: true,
         entrypoint: ep.id,
-        message: `${ep.label} accesses ${tables.join(', ')} by a request-supplied id with no ownership scope proven on the path — verify object-level authorization (BOLA/IDOR)`,
+        ownership,
+        message:
+          `${ep.label} accesses ${tables.join(', ')} by a request-supplied id with no ownership scope proven on the path` +
+          (hints.length ? ` — ${hints.join('; ')}` : '') +
+          ` — verify object-level authorization (BOLA/IDOR)`,
         evidence: idScoped.map((n) => `${n.id} (${locOf(n)})`),
       });
     }
@@ -251,7 +276,55 @@ export function checkGraph(graph) {
     polarity.push({ entrypoint: ep.id, vector: vec });
   }
 
-  return { findings: sortFindings(findings), obligations, polarity };
+  // Lateral inhibition (ADR-060): a rule that fires on a large FRACTION of the routes is a
+  // codebase-wide PATTERN, not N independent findings — 97 identical lines bury the rare, sharp
+  // signals (loss of contrast). Collapse a flooding rule into ONE summary at its MAX severity.
+  // SOUND + verdict-neutral: a hard rule stays hard and still gates (we never HIDE a finding — a
+  // suppressed danger would be a false PROVEN); we only stop it flooding the report. The retina
+  // suppresses uniform light, never an edge.
+  const collapsed = collapseFloods(findings, g.entrypoints.length);
+
+  return { findings: sortFindings(collapsed), obligations, polarity };
+}
+
+// A rule is "pervasive" when it fires on more than this fraction of the routes AND on at least
+// FLOOD_MIN of them — measured on the corpus: real floods sit at 18% (dub member-bypass) / 41%
+// (directus irreversible) on 100+ routes, while the sharp per-route signals (BOLA 10%, unvalidated
+// 11%, unguarded 1%) stay below. The absolute floor keeps a small app (a 2-route fixture where one
+// finding is trivially "50%") from ever collapsing — a pattern needs real breadth to be a pattern.
+const FLOOD_DENSITY = 0.15;
+const FLOOD_MIN = 10;
+export function collapseFloods(findings, totalRoutes) {
+  if (!totalRoutes) return findings;
+  const byRule = new Map();
+  for (const f of findings) {
+    if (!byRule.has(f.rule)) byRule.set(f.rule, []);
+    byRule.get(f.rule).push(f);
+  }
+  const out = [];
+  for (const [rule, list] of [...byRule].sort((a, b) => cmp(a[0], b[0]))) {
+    const routes = new Set(list.map((f) => f.entrypoint));
+    if (routes.size >= FLOOD_MIN && routes.size / totalRoutes > FLOOD_DENSITY) {
+      // strongest severity wins; advisory only if EVERY collapsed finding was advisory (a hard
+      // finding in the flood keeps the summary hard — it must still gate the verdict).
+      const severity = list
+        .map((f) => f.severity)
+        .sort((a, b) => SEVERITY_RANK[a] - SEVERITY_RANK[b])[0];
+      const anyHard = list.some((f) => !f.advisory);
+      out.push({
+        rule,
+        severity,
+        ...(anyHard ? {} : { advisory: true }),
+        entrypoint: '(codebase-wide)',
+        pervasive: routes.size,
+        message: `${rule} is pervasive — fires on ${routes.size}/${totalRoutes} routes; review the pattern, not each route (all routes listed in evidence)`,
+        evidence: [...routes].sort(),
+      });
+    } else {
+      out.push(...list);
+    }
+  }
+  return out;
 }
 
 // ---------------------------------------------------------------------------
@@ -365,6 +438,38 @@ export function countProvable(graph) {
   return n;
 }
 
+// Ownership-model inference (BolaRay, CCS 2024): from a resource table's declared columns and
+// foreign keys, which object-level authorization model SHOULD it carry? DIRECT_OWNER (a
+// user/owner column), GROUP_SCOPED (a workspace/team/tenant column — caller must belong to the
+// group), TRANSITIVE (ownership reached via a FK to an owned table), or null (a shared/global
+// resource, or intent SPARDA cannot see). Columns are lowercased (prisma.js / sql.js). Pure and
+// bounded. Used ONLY to enrich the O7 advisory — it never changes the verdict (soundness: the
+// schema reveals the model, never the runtime authorization intent, which is the semantic gap
+// OWASP/BolaRay both name as the reason no static tool can PROVE access control).
+const OWNER_DIRECT = /^(user|owner|author|creator)_?id$/;
+const OWNER_GROUP =
+  /^(workspace|project|team|tenant|org|organization|store|customer|partner|program|account|company|group)_?id$/;
+function inferOwnershipModel(tableName, statesByTable, seen = new Set(), depth = 0) {
+  const st = statesByTable.get(tableName);
+  if (!st || seen.has(tableName) || depth > 3) return null;
+  seen.add(tableName);
+  const cols = (st.meta?.columns ?? []).map((c) => c.name);
+  const direct = cols.find((n) => OWNER_DIRECT.test(n));
+  if (direct) return { model: 'direct-owner', key: direct };
+  const group = cols.find((n) => OWNER_GROUP.test(n));
+  if (group) return { model: 'group-scoped', key: group };
+  for (const ref of st.meta?.references ?? []) {
+    if (!ref.table || ref.table === tableName) continue;
+    const sub = inferOwnershipModel(ref.table, statesByTable, seen, depth + 1);
+    if (sub)
+      return {
+        model: 'transitive',
+        key: `${(ref.fields ?? []).join(',')}->${ref.table}.${sub.key}`,
+      };
+  }
+  return null;
+}
+
 // Below this resolved/surface coverage ratio, a CLEAN app cannot claim PROVEN — it
 // resolved too little of its own behavior for "no violation" to mean anything (the
 // PROVEN-COMPLETE vs PROVEN-PARTIAL line). Measured floor: real PROVEN apps sit at 60%+

package/src/ubg/llm-resolve.js

@@ -0,0 +1,46 @@
+// ubg/llm-resolve.js — LLM-assisted resolution of unverified guards, with the
+// verify-before-admit guardrail the literature identifies as the critical step (Thakur et al.,
+// "Interleaving static analysis and LLM prompting", STTT 2025; "Boosting Pointer Analysis with
+// LLM-Enhanced Allocation Detection", arXiv 2509.22530). See docs/RESEARCH-AND-10X-IDEAS §Part 1.
+//
+// THE CONTRACT (soundness — non-negotiable): an LLM may only produce a *resolution hint* (where a
+// guard's deny logic likely lives) or a *candidate classification*. It never asserts behavior.
+// SPARDA re-verifies the hint STRUCTURALLY against the real graph before anything enters the
+// proof. A guard becomes `verified` ONLY when a deny path is structurally proven — never on the
+// model's word. Skipping this exact step is what produced E-022/E-025/E-026 (false PROVEN), the
+// cardinal sin. This module makes the guardrail executable and testable; the hint *producer*
+// (MCP sampling — the host never pays, CLAUDE.md rule 1) plugs in on top and is out of scope
+// here on purpose, so no LLM dependency and no un-verified path can ever leak into the graph.
+
+// The unverified guards worth an LLM resolution attempt: protection asserted by name, whose body
+// SPARDA never saw deny (the `unverified-guard` blind-spot category, blindspots.js). These are
+// the ambiguous DI/factory/opaque-import patterns the AST walker can't resolve structurally.
+export function resolutionTargets(graph) {
+  if (!graph?.nodes) return [];
+  return graph.nodes.filter((n) => n.kind === 'guard' && !n.meta?.verified);
+}
+
+// Admit an LLM resolution hint for ONE guard — and ONLY if SPARDA's own structural prover
+// confirms a deny path at the hinted location. `proveDeny(hint)` is SPARDA's deny-detection,
+// injected so this module stays testable and carries no LLM dependency itself. The hint is NEVER
+// trusted on its own; it only tells the structural prover where to look.
+//
+// On success the guard is marked `verified: true` with `verifiedVia: 'llm-guided'` — a distinct
+// provenance so an LLM-guided verification is always auditable and never silently indistinguish-
+// able from a natively-proven guard (guardsVerified counting, dossier, audits can separate them).
+// Returns { admitted, reason }; mutates guardNode.meta only on a confirmed admit.
+export function admitResolutionHint(guardNode, hint, { proveDeny } = {}) {
+  if (!guardNode || guardNode.kind !== 'guard')
+    return { admitted: false, reason: 'not-a-guard' };
+  if (guardNode.meta?.verified) return { admitted: false, reason: 'already-verified' };
+  if (typeof proveDeny !== 'function')
+    return { admitted: false, reason: 'no-structural-prover' };
+  if (!hint) return { admitted: false, reason: 'no-hint' };
+
+  // THE GUARDRAIL. The model's hint buys nothing until SPARDA itself proves the deny.
+  const denies = proveDeny(hint) === true;
+  if (!denies) return { admitted: false, reason: 'hint-not-structurally-confirmed' };
+
+  guardNode.meta = { ...guardNode.meta, verified: true, verifiedVia: 'llm-guided' };
+  return { admitted: true, reason: 'structurally-confirmed' };
+}

package/src/ubg/prisma.js

@@ -12,6 +12,48 @@ import path from 'node:path';
 import { cmp } from './schema.js';
 
 const SCHEMA_CANDIDATES = ['prisma/schema.prisma', 'schema.prisma', 'db/schema.prisma'];
+// Prisma's `prismaSchemaFolder` layout (stable since 6.x): the schema is SPLIT across many
+// `*.prisma` files under a directory, no single schema.prisma. Modern apps use it (dub: 36
+// files). Without this, their entire state layer — invariants, aggregates, ownership — is
+// invisible (E-046). These directory candidates are scanned when no single file is found.
+const SCHEMA_DIR_CANDIDATES = ['prisma/schema', 'schema', 'db/schema'];
+
+// Gather every `.prisma` file to parse: the single file if present, else all files in a schema
+// folder (bounded, deterministic order). Returns [{ file, rel }] or [] if nothing found.
+function collectSchemaFiles(cwd, fileCandidates, dirCandidates) {
+  for (const c of fileCandidates) {
+    const abs = path.resolve(cwd, c);
+    if (fs.existsSync(abs) && fs.statSync(abs).isFile()) return [abs];
+  }
+  for (const c of dirCandidates) {
+    const abs = path.resolve(cwd, c);
+    let stat;
+    try {
+      stat = fs.statSync(abs);
+    } catch {
+      continue;
+    }
+    if (!stat.isDirectory()) continue;
+    const files = [];
+    const walk = (dir, depth) => {
+      if (depth > 3) return;
+      let entries;
+      try {
+        entries = fs.readdirSync(dir, { withFileTypes: true });
+      } catch {
+        return;
+      }
+      for (const e of entries.sort((a, b) => (a.name < b.name ? -1 : 1))) {
+        const p = path.join(dir, e.name);
+        if (e.isDirectory()) walk(p, depth + 1);
+        else if (e.name.endsWith('.prisma')) files.push(p);
+      }
+    };
+    walk(abs, 0);
+    if (files.length) return files;
+  }
+  return [];
+}
 
 const TYPE_MAP = {
   string: 'string',
@@ -36,42 +78,47 @@ export function parsePrismaSchemas(cwd) {
   } catch {
     // no package.json / unparsable — the default candidates stand
   }
-  const file = candidates
-    .map((c) => path.resolve(cwd, c))
-    .find((abs) => fs.existsSync(abs));
-  if (!file) return { tables: [], skipped, sourceFile: null };
-
-  let src;
-  try {
-    src = fs.readFileSync(file, 'utf8');
-  } catch (err) {
-    skipped.push({ reason: `unreadable schema.prisma: ${err.message}` });
-    return { tables: [], skipped, sourceFile: null };
+  const files = collectSchemaFiles(cwd, candidates, SCHEMA_DIR_CANDIDATES);
+  if (!files.length) return { tables: [], skipped, sourceFile: null };
+
+  // Read + strip comments once per file; keep the relative path for correct per-table locs.
+  const parts = [];
+  for (const abs of files) {
+    try {
+      parts.push({
+        rel: path.relative(cwd, abs).split(path.sep).join('/'),
+        clean: fs.readFileSync(abs, 'utf8').replace(/\/\/[^\n]*/g, ''),
+      });
+    } catch (err) {
+      skipped.push({ reason: `unreadable prisma schema: ${err.message}` });
+    }
   }
-  const sourceFile = path.relative(cwd, file).split(path.sep).join('/');
-  const clean = src.replace(/\/\/[^\n]*/g, '');
+  if (!parts.length) return { tables: [], skipped, sourceFile: null };
+  const sourceFile = parts.length === 1 ? parts[0].rel : path.dirname(parts[0].rel);
 
-  // enums first — model fields reference them
+  // Enums and model names are collected ACROSS ALL files first — a model in one file may
+  // reference an enum (or a relation target) declared in another (the whole point of the
+  // folder layout). Then models are parsed per-file so each table keeps its true file:line.
   const enums = new Map(); // lowercase name -> [values lowercase]
-  for (const m of clean.matchAll(/enum\s+(\w+)\s*\{([^}]*)\}/g)) {
-    const values = m[2]
-      .split(/\s+/)
-      .map((v) => v.trim())
-      .filter((v) => v && !v.startsWith('@'))
-      .map((v) => v.toLowerCase());
-    enums.set(m[1].toLowerCase(), values);
+  const modelNames = new Set();
+  for (const { clean } of parts) {
+    for (const m of clean.matchAll(/enum\s+(\w+)\s*\{([^}]*)\}/g)) {
+      const values = m[2]
+        .split(/\s+/)
+        .map((v) => v.trim())
+        .filter((v) => v && !v.startsWith('@'))
+        .map((v) => v.toLowerCase());
+      enums.set(m[1].toLowerCase(), values);
+    }
+    for (const m of clean.matchAll(/model\s+(\w+)\s*\{/g)) modelNames.add(m[1]);
   }
 
-  const modelNames = new Set([...clean.matchAll(/model\s+(\w+)\s*\{/g)].map((m) => m[1]));
-
   const tables = [];
-  for (const m of clean.matchAll(/model\s+(\w+)\s*\{([^}]*)\}/g)) {
-    const modelName = m[1];
-    const body = m[2];
-    const line = clean.slice(0, m.index).split('\n').length;
-    tables.push(
-      parseModel(modelName, body, line, sourceFile, enums, modelNames, skipped),
-    );
+  for (const { rel, clean } of parts) {
+    for (const m of clean.matchAll(/model\s+(\w+)\s*\{([^}]*)\}/g)) {
+      const line = clean.slice(0, m.index).split('\n').length;
+      tables.push(parseModel(m[1], m[2], line, rel, enums, modelNames, skipped));
+    }
   }
   tables.sort((a, b) => a.name.localeCompare(b.name));
   return { tables, skipped, sourceFile };

package/src/ubg/stitch.js

@@ -0,0 +1,105 @@
+// ubg/stitch.js — cross-repo / cross-service stitching (the moat). SAST incumbents
+// (CodeQL/Semgrep/Snyk) are mono-repo by construction — a violation that crosses a service
+// boundary (service A forwards a caller-supplied id to service B, which trusts it blindly) is
+// invisible to all of them. SPARDA already produces the artifact that makes it possible: a
+// committed `ubg.json` per repo. This pass reads N of them and joins the outbound HTTP calls of
+// one to the entrypoints of another — the reproduction of BACTERIAL QUORUM SENSING: each service
+// emits a signal (its graph); the collective behavior emerges from reading them together, with no
+// central coordinator and no network access between services (each graph is committed in git).
+//
+// Soundness: cross-service findings are ADVISORY (same stance as BOLA/O7) — the join is a
+// structural match, never a proof of runtime intent across the boundary. It points a human at a
+// trust boundary to review; it never gates a verdict.
+
+// A route path or an outbound URL target → comparable segment list. Strips scheme+host and query,
+// lowercases, collapses every parameter form (:id, {id}, *) to a single '*' token.
+function segmentsOf(target) {
+  let s = String(target ?? '');
+  const url = s.match(/^[a-z][a-z0-9+.-]*:\/\/[^/]+(\/.*)?$/i);
+  if (url) s = url[1] ?? '/';
+  s = s.split(/[?#]/)[0].toLowerCase();
+  return s
+    .split('/')
+    .filter(Boolean)
+    .map((seg) => (/^[:{*]/.test(seg) || seg.includes('*') ? '*' : seg));
+}
+
+// method from an entrypoint label ("GET /users/:id") or an http_call ("httpMethod"); default GET.
+const methodOf = (label) =>
+  (String(label).match(/^([A-Z]+)\s/)?.[1] ?? 'GET').toUpperCase();
+const pathPart = (label) => String(label).replace(/^[A-Z]+\s+/, '');
+
+// B's entrypoint (callee) matches A's outbound call when the methods agree and the callee's path
+// segments are a SUFFIX of the caller's target segments — so a base-URL prefix
+// (`http://host/api/v1/users/*` calling B's `/users/:id`) still matches. '*' matches anything.
+function suffixMatch(callerSegs, calleeSegs) {
+  if (calleeSegs.length === 0 || calleeSegs.length > callerSegs.length) return false;
+  const off = callerSegs.length - calleeSegs.length;
+  for (let i = 0; i < calleeSegs.length; i++) {
+    const a = callerSegs[off + i];
+    const b = calleeSegs[i];
+    if (a !== b && a !== '*' && b !== '*') return false;
+  }
+  return true;
+}
+
+// services: [{ name, graph (canonical), findings? }]. Returns cross-service call edges and the
+// advisory findings that ride them.
+export function stitchServices(services) {
+  const edges = [];
+  const findings = [];
+
+  // index every service's entrypoints (the callees) by method + segments
+  const callees = [];
+  for (const svc of services) {
+    for (const n of svc.graph?.nodes ?? []) {
+      if (n.kind !== 'entrypoint') continue;
+      callees.push({
+        service: svc.name,
+        entrypoint: n.id,
+        label: n.label,
+        method: methodOf(n.label),
+        segs: segmentsOf(pathPart(n.label)),
+        bola: (svc.findings ?? []).some(
+          (f) => f.rule === 'OBJECT_SCOPE_UNPROVEN' && f.entrypoint === n.id,
+        ),
+      });
+    }
+  }
+
+  for (const svc of services) {
+    for (const n of svc.graph?.nodes ?? []) {
+      if (n.kind !== 'effect' || n.meta?.effectType !== 'http_call') continue;
+      const callerMethod = (n.meta.httpMethod ?? 'GET').toUpperCase();
+      const callerSegs = segmentsOf(n.meta.target);
+      if (!callerSegs.length) continue;
+      for (const c of callees) {
+        if (c.service === svc.name) continue; // never stitch a service to itself
+        if (c.method !== callerMethod) continue;
+        if (!suffixMatch(callerSegs, c.segs)) continue;
+        edges.push({
+          fromService: svc.name,
+          fromEffect: n.id,
+          toService: c.service,
+          toEntrypoint: c.entrypoint,
+          method: callerMethod,
+          path: '/' + c.segs.join('/'),
+        });
+        // the trust-boundary finding: A forwards to a B endpoint B itself flagged as accessing an
+        // object by a request id with no ownership scope — so the object-level authorization
+        // across the A→B boundary is unproven. Advisory (a structural match, not runtime intent).
+        if (c.bola) {
+          findings.push({
+            rule: 'CROSS_SERVICE_OBJECT_SCOPE',
+            severity: 'info',
+            advisory: true,
+            entrypoint: c.entrypoint,
+            message: `${svc.name} calls ${c.service} ${c.method} ${'/' + c.segs.join('/')}, which accesses an object by a request id with no ownership scope proven — the object-level authorization across this service boundary is unproven (cross-service BOLA/IDOR)`,
+            evidence: [`${svc.name}:${n.id}`, `${c.service}:${c.entrypoint}`],
+          });
+        }
+      }
+    }
+  }
+  return { edges, findings };
+}