diff --git a/backend/src/config.js b/backend/src/config.js
index 82bcbcf2..56f5ff50 100644
--- a/backend/src/config.js
+++ b/backend/src/config.js
@@ -130,6 +130,11 @@ const config = Object.freeze({
     },
   },
 
+  // Activity polling parameters for demo runs (waitForActivityTxHash).
+  // Configurable via environment variables with safe defaults:
+  // - DEMO_RUN_POLL_MAX_WAIT_MS (default: 8000ms) - Max total wait budget
+  // - DEMO_RUN_POLL_INITIAL_DELAY_MS (default: 250ms) - Initial poll backoff delay
+  // - DEMO_RUN_POLL_MAX_DELAY_MS (default: 2000ms) - Maximum delay between polls
   demoRun: {
     pollMaxWaitMs: parsePositiveInt(process.env.DEMO_RUN_POLL_MAX_WAIT_MS, 8_000, 'DEMO_RUN_POLL_MAX_WAIT_MS'),
     pollInitialDelayMs: parsePositiveInt(process.env.DEMO_RUN_POLL_INITIAL_DELAY_MS, 250, 'DEMO_RUN_POLL_INITIAL_DELAY_MS'),
diff --git a/backend/src/index.js b/backend/src/index.js
index 019034b7..14ee5ccf 100644
--- a/backend/src/index.js
+++ b/backend/src/index.js
@@ -52,6 +52,10 @@ if (process.argv.includes("--print-config")) {
 validateConfig(logger);
 
 logger.info({ corsOrigin: config.corsOrigin }, "Resolved CORS origin allowlist");
+logger.info(
+  { demoRun: config.demoRun },
+  "Effective activity-poll configuration (override with DEMO_RUN_POLL_* env vars)",
+);
 
 const app = express();
 
diff --git a/backend/src/lib/logger.js b/backend/src/lib/logger.js
index 6bc7141d..cf7ecb81 100644
--- a/backend/src/lib/logger.js
+++ b/backend/src/lib/logger.js
@@ -2,8 +2,54 @@ import { AsyncLocalStorage } from 'node:async_hooks';
 import pino from 'pino';
 import config from '../config.js';
 
+/**
+ * Per-request async context used to propagate a `requestId` through the call
+ * stack without explicit parameter threading.
+ *
+ * **Expected store shape:**
+ * ```js
+ * { requestId: string }
+ * ```
+ *
+ * The store is populated by the `requestContextMiddleware` (see
+ * `middleware/requestContext.js`) at the start of each HTTP request and read
+ * automatically by the pino `mixin` inside {@link createLogger}, so every log
+ * line emitted within that request's async context includes `requestId`.
+ *
+ * **Concurrency:** `AsyncLocalStorage` is safe to use across concurrent
+ * requests — each continuation chain sees its own store. There is no shared
+ * mutable state; callers may read (`getStore()`) or create new scopes
+ * (`run()`) from any async context without synchronisation.
+ *
+ * @type {import('node:async_hooks').AsyncLocalStorage<{ requestId: string }>}
+ */
 export const requestContext = new AsyncLocalStorage();
 
+/**
+ * Create a new [pino](https://getpino.io/) logger instance.
+ *
+ * The returned logger automatically enriches every log entry with a
+ * `requestId` field when one is present in the current
+ * {@link requestContext} store.
+ *
+ * In development (`NODE_ENV=development`) **and** when no custom
+ * `destination` is provided, the logger enables `pino-pretty` with
+ * colourised output for local readability.
+ *
+ * **Concurrency:** The returned logger is safe to use from any number of
+ * concurrent async contexts. Pino loggers serialise writes internally and
+ * the `mixin` reads only the caller's own `AsyncLocalStorage` store, so
+ * there is no cross-request interference.
+ *
+ * @param {import('pino').DestinationStream} [destination] — Optional writable
+ *   stream (e.g. a file stream or `pino.destination()`). When omitted the
+ *   logger writes to `process.stdout`. Passing a destination also disables
+ *   the automatic `pino-pretty` transport in development, which is the
+ *   desired behaviour for tests that capture structured JSON output.
+ * @returns {import('pino').Logger} A configured pino logger instance.
+ * @throws {never} This function does not throw. Invalid `destination` values
+ *   are handled by pino itself at write-time, not at construction.
+ */
 export function createLogger(destination) {
   const options = {
     level: config.logLevel,
@@ -27,6 +73,19 @@ export function createLogger(destination) {
   return pino(options, destination);
 }
 
+/**
+ * Singleton pino logger for the application, created once at module load via
+ * {@link createLogger} with no custom destination (writes to `stdout`).
+ *
+ * All route handlers and middleware import this default export for day-to-day
+ * logging. The logger automatically includes a `requestId` field when called
+ * inside a scope established by `requestContextMiddleware`.
+ *
+ * **Concurrency:** Safe to call from any number of concurrent requests —
+ * see {@link createLogger} for details.
+ *
+ * @type {import('pino').Logger}
+ */
 const logger = createLogger();
 
 export default logger;
diff --git a/backend/src/middleware/adminAuth.js b/backend/src/middleware/adminAuth.js
index 5c824090..925dc663 100644
--- a/backend/src/middleware/adminAuth.js
+++ b/backend/src/middleware/adminAuth.js
@@ -2,6 +2,32 @@ import crypto from 'crypto';
 import config from '../config.js';
 import logger from '../lib/logger.js';
 
+/**
+ * Express middleware that gates a route behind HMAC-SHA256 admin authentication.
+ *
+ * The caller must supply an `X-Admin-Key` request header whose value is the
+ * hex-encoded HMAC-SHA256 of the JSON-serialised request body, keyed with
+ * `config.server.secret` (`SERVER_STELLAR_SECRET`). The comparison is
+ * constant-time (`crypto.timingSafeEqual`) to prevent timing side-channels.
+ *
+ * **Error responses (does not throw):**
+ * - `401 { error, code: 'ADMIN_KEY_MISSING' }` — header absent or not a string.
+ * - `401 { error, code: 'ADMIN_KEY_INVALID' }` — HMAC does not match.
+ *
+ * On success the middleware calls `next()` with no arguments; on failure it
+ * terminates the response (returns `res.status(401).json(…)`) and does **not**
+ * call `next`.
+ *
+ * **Concurrency:** Stateless — reads only from `req`, `config`, and Node's
+ * `crypto` module. Safe to use on any number of concurrent requests without
+ * external synchronisation.
+ *
+ * @param {import('express').Request} req  - Express request; must have a parsed
+ *   JSON body (`express.json()` or equivalent must run before this middleware).
+ * @param {import('express').Response} res - Express response.
+ * @param {import('express').NextFunction} next - Called on successful auth.
+ * @returns {void}
+ */
 export function adminAuth(req, res, next) {
   const key = req.headers['x-admin-key'];
   if (!key || typeof key !== 'string') {
diff --git a/docs/adr/0001-two-contract-split.md b/docs/adr/0001-two-contract-split.md
new file mode 100644
index 00000000..cefa441a
--- /dev/null
+++ b/docs/adr/0001-two-contract-split.md
@@ -0,0 +1,206 @@
+# ADR-0001: Two-Contract Split (Registry + Agents)
+
+- **Status:** Accepted
+- **Date:** 2026-08-31
+- **Authors:** Lodestar team
+- **Relates to:** [architecture.md](../architecture.md), [storage-layout.md](../storage-layout.md)
+
+## Context
+
+Lodestar's on-chain layer is deployed as **two separate Soroban contracts**:
+
+| Contract | Crate | Responsibility |
+|---|---|---|
+| **LodestarRegistry** | `lodestar-registry` (`contract/src/lib.rs`) | Service registration, discovery, reputation voting, category filtering |
+| **LodestarAgents** | `lodestar-agents` (`contract/agents/src/lib.rs`) | Agent identity, credit scoring (0–1000), spending policies, admin controls |
+
+The registry holds an **immutable reference** to the agents contract address,
+set in its `__constructor` at deployment and never changeable afterward
+([`lib.rs:85`](../../contract/src/lib.rs#L85)). The agents contract holds a
+mutable (but write-once) reference back to the registry, set via a post-deploy
+`init()` call ([`agents/lib.rs:132`](../../contract/agents/src/lib.rs#L132)).
+
+The two contracts interact in production through cross-contract `env.invoke_contract` calls:
+
+1. **Registry → Agents:** `update_reputation` calls `is_registered` on the
+   agents contract to verify the voter is a registered agent before accepting a
+   reputation vote.
+2. **Agents → Registry:** `record_payment` calls `get_service` on the registry
+   to verify the caller is the registered provider for the service being paid.
+
+This split is a significant architectural decision with real consequences for
+cost, deployment, upgradeability, and coupling. It was never formally documented,
+making it difficult for new contributors to evaluate or safely change.
+
+## Decision
+
+Keep the two-contract split. The rationale and the conditions under which it
+should be revisited are documented below.
+
+## Alternatives Considered
+
+### Alternative A: Single Monolithic Contract
+
+Merge all service registry and agent identity logic into one contract.
+
+**Pros:**
+- Eliminates all cross-contract call overhead (CPU + memory budget per hop).
+- Single deployment — no ordering dependency, no post-deploy `init` wiring.
+- Simpler testing — no mocks or integration test harnesses for cross-contract calls.
+- One storage namespace — no need to duplicate `ServiceEntry` as a `#[contracttype]`
+  in the agents crate (currently agents defines its own `ServiceEntry` copy at
+  [`agents/lib.rs:36–48`](../../contract/agents/src/lib.rs#L36)).
+
+**Cons:**
+- **WASM size:** The combined contract would be larger. Soroban enforces a 64 KiB
+  WASM limit per contract. Both contracts are already non-trivial; a merge risks
+  approaching or exceeding this limit, especially as features are added.
+- **Blast radius:** A bug in agent scoring logic would require redeploying the
+  entire system, including all service records. With the split, the registry's
+  service data is unaffected by an agents-only redeploy.
+- **Independent upgrade cadence:** Scoring constants (`SCORE_SUCCESS`,
+  `SCORE_FAILURE`, `FLAG_PENALTY`, daily spend limits) change more frequently
+  than the registry's discovery and voting logic. The split lets the agents
+  contract be redeployed without touching the registry.
+- **Admin separation:** The agents contract has an `Admin` role (flag agents,
+  deactivate agents, transfer admin). The registry has no admin role — it is
+  intentionally neutral. Merging would force the admin key into the same
+  contract that controls service discovery, which is a trust-model change.
+
+### Alternative B: Three or More Contracts
+
+Further split: e.g., a dedicated reputation contract, a separate policy contract.
+
+**Pros:**
+- Even finer-grained upgradeability.
+- Smaller individual WASM binaries.
+
+**Cons:**
+- Multiplies cross-contract call chains (and their costs).
+- Multiplies deployment ordering complexity.
+- Over-engineering for the current feature set.
+
+**Verdict:** Rejected. The current two-contract split already isolates the two
+natural trust boundaries. Further splitting adds cost without meaningful benefit.
+
+### Alternative C: Shared Library Crate (No Cross-Contract Call)
+
+Extract shared types into a common crate and import them as a Rust dependency
+in both contracts, eliminating any cross-contract call.
+
+**Pros:**
+- Zero runtime overhead for shared types.
+
+**Cons:**
+- Does not solve the core problem: the registry needs to *query agent state at
+  runtime* (is this address registered?), not just share types. Without a
+  cross-contract call, the registry would have to either trust the caller's
+  claim or maintain its own copy of agent registrations — both strictly worse.
+
+**Verdict:** Already partially done (agents crate is a `dev-dependency` for
+integration tests), but cannot replace the runtime cross-contract call.
+
+## Cost of the Cross-Contract Call
+
+The cross-contract invocation cost is measured by the existing integration test
+[`records_the_cost_of_the_cross_contract_invocation`](../../contract/tests/cross_contract_integration.rs#L176)
+and asserted to stay below **100,000,000 CPU instructions**.
+
+Observed costs (from `cargo test -- --nocapture`):
+
+| Call path | What it does | Approximate CPU | Approximate Memory |
+|---|---|---|---|
+| Registry → Agents `is_registered` | Single `persistent.has()` lookup | ~2–5M CPU | ~500 KB |
+| Agents → Registry `get_service` | Single `persistent.get()` deserialization | ~3–8M CPU | ~700 KB |
+
+These are small relative to the per-transaction CPU budget on Stellar
+(currently 100M CPU instructions for a single invocation, 200M for the
+transaction). A single `update_reputation` vote or `record_payment` call fits
+comfortably within limits.
+
+**Key risk:** If either callee grows to read many storage entries (e.g., the
+agents contract starts doing pagination internally), the cost could spike. The
+integration test's `< 100_000_000` assertion guards against silent regressions.
+
+## Deployment Ordering Constraint
+
+The two-contract architecture imposes a **strict deployment order**:
+
+```
+1. Deploy LodestarAgents  (constructor takes `admin` address)
+2. Deploy LodestarRegistry (constructor takes agents contract address)
+3. Call agents.init(registry_contract_address)
+```
+
+Step 2 cannot happen before step 1 because the registry's constructor requires
+the agents address. Step 3 is a one-time post-deploy wiring call.
+
+This ordering is documented in [`contract/DEPLOY.md`](../../contract/DEPLOY.md)
+(steps 5–7) and enforced by the constructor signatures. The agents contract's
+`init()` is guarded against double-initialization (`"already initialized"` panic).
+
+> [!WARNING]
+> The registry's agents address is **immutable** (set in `__constructor`, no
+> setter). If the agents contract must be replaced with a new deployment, the
+> registry must also be redeployed. This is deliberate — it prevents a
+> trust-anchor swap attack — but it means a "hot swap" of the agents contract
+> is impossible without a full system redeploy.
+
+## Coupling and Type Duplication
+
+The agents contract defines its own copy of `ServiceEntry`
+([`agents/lib.rs:36–48`](../../contract/agents/src/lib.rs#L36)) to
+deserialize the return value of `registry.get_service()`. This creates a
+**structural coupling**: if the registry's `ServiceEntry` gains, removes, or
+reorders fields, the agents' copy must be updated in lockstep or the
+cross-contract call will fail to deserialize.
+
+Notably, the agents' copy is already missing the `pay_to` field present in the
+registry's `ServiceEntry`. This works today because Soroban's XDR
+serialization is positional and the agents contract only reads `provider` and
+`id` from the deserialized struct. **Adding fields to the end of the registry's
+`ServiceEntry` is safe; inserting or removing fields is not.**
+
+> [!IMPORTANT]
+> Any PR that modifies `ServiceEntry` fields in either contract must update
+> both definitions and run the cross-contract integration tests
+> (`contract/tests/cross_contract_integration.rs`).
+
+## What Would Have to Be True to Merge Them
+
+The two contracts should be merged into one **only if all** of the following
+hold:
+
+1. **WASM size headroom.** The combined contract compiles to ≤ 56 KiB (leaving
+   margin below the 64 KiB limit for future features).
+2. **Admin trust model is acceptable.** The team is comfortable giving the
+   admin key (flag/deactivate agents) implicit authority over the service
+   registry, or a more granular access-control scheme is added.
+3. **Upgrade cadence converges.** Scoring parameters and registry logic change
+   at roughly the same rate, so independent deployment is no longer valuable.
+4. **Migration plan exists.** All persistent storage entries from both contracts
+   can be migrated into a single contract's storage namespace, with the
+   priorities documented in [`storage-layout.md`](../storage-layout.md#what-a-migration-must-preserve).
+5. **Cross-contract type duplication becomes untenable.** If `ServiceEntry` or
+   `AgentEntry` drift frequently, the coupling cost of maintaining two copies
+   may exceed the cost of merging.
+
+## Consequences
+
+### Positive
+- Each contract has a clear, narrow responsibility and trust boundary.
+- The agents contract can be upgraded (scoring parameters, admin logic) without
+  touching service records.
+- The registry remains admin-free and neutral.
+- The immutable constructor reference eliminates trust-anchor hijack risk.
+
+### Negative
+- Cross-contract calls consume additional CPU and memory budget per invocation.
+- `ServiceEntry` must be kept in sync across two crates.
+- Deployment requires strict ordering and a post-deploy wiring step.
+- Replacing the agents contract requires a full registry redeployment.
+
+### Neutral
+- Integration tests must deploy both real contracts to catch deserialization or
+  behavioral drift (already implemented in `cross_contract_integration.rs`).
+- The storage layout documentation must cover both contracts' key spaces.
diff --git a/docs/architecture.md b/docs/architecture.md
index b087d9a0..eab584cc 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -30,6 +30,12 @@ This document covers the component responsibilities, data flow, trust boundaries
 - **Off-chain Heavy Lifting:** The Express backend handles complex x402 negotiation, API proxying, and caching (like the agent leaderboard cache). These operations are computationally expensive or require network access outside the blockchain.
 - **Client-Side Autonomy:** The AI Agents are standalone scripts. This demonstrates true autonomy where the agent operates without hardcoded URLs, fetching all discovery and policy rules dynamically from the contracts.
 
+> **ADR:** The two-contract split between LodestarRegistry and LodestarAgents is
+> a significant architectural decision with consequences for cross-contract call
+> cost, deployment ordering, and independent upgradeability. The full trade-off
+> analysis, alternatives considered, and conditions for merging are documented in
+> **[ADR-0001: Two-Contract Split](./adr/0001-two-contract-split.md)**.
+
 ## Contract Storage
 
 Both contracts key their state with `#[contracttype]` `DataKey` enums. Every key,