@aztec/stdlib 0.0.1-commit.993d52e → 0.0.1-commit.9a89641

package/src/deserialization/index.ts

@@ -11,8 +11,41 @@ export const MAX_TXS_PER_BLOCK = 2 ** 16;
 /** Max committee size - theoretical max from bitmap design (256 bytes × 8 bits) */
 export const MAX_COMMITTEE_SIZE = 2048;
 
-/** Max blocks per checkpoint */
-export const MAX_BLOCKS_PER_CHECKPOINT = 72;
+/**
+ * Maximum number of L2 blocks in a provable checkpoint. Used for deserialization and when ingesting
+ * checkpoints already accepted by L1.
+ *
+ * This MUST be >= the number of blocks the proving system can carry, or a structurally-valid, provable
+ * checkpoint that L1 accepts would be rejected on ingest (`validateCheckpoint`) and wedge the archiver.
+ * The circuits and L1 impose no explicit per-checkpoint block count; the ceiling comes from the
+ * blob-field budget:
+ *
+ *   budget = BLOBS_PER_CHECKPOINT * FIELDS_PER_BLOB = 6 * 4096 = 24,576 fields
+ *
+ *   per-block minimal blob footprint:
+ *     first block  (may be empty)   = 7 fields   (NUM_FIRST_BLOCK_END_BLOB_FIELDS)
+ *     every other  (needs >= 1 tx)  = 6 + 4 = 10 fields
+ *                                     (NUM_BLOCK_END_BLOB_FIELDS + a 4-field minimal tx:
+ *                                      tx start marker + tx hash + fee + 1 mandatory nullifier)
+ *     + 1 checkpoint-end marker      (NUM_CHECKPOINT_END_MARKER_FIELDS)
+ *
+ *   max N:  7 + 10*(N - 1) + 1 <= 24,576  =>  10*(N - 1) <= 24,568  =>  N <= 2,457
+ *
+ * Only the first block may be empty; every other block needs >= 1 tx (the circuits have no empty-tx
+ * variant for non-first blocks), so 2457 is the largest provable checkpoint. The blob format can encode
+ * more blocks than this (up to ~4095 with all-empty blocks), but such a checkpoint is unprovable and
+ * can only reach L1 with a malicious committee supermajority — a terminal network compromise where a
+ * wedged archiver is an acceptable outcome. We therefore bound ingest to the provable maximum.
+ * Invariant checked by the unit test in deserialization.test.ts.
+ */
+export const MAX_CAPACITY_BLOCKS_PER_CHECKPOINT = 2457;
+
+/**
+ * Max blocks per checkpoint we are willing to build or attest to (conservative client policy, and the
+ * default block-count limit for `validateCheckpoint`). Distinct from {@link MAX_CAPACITY_BLOCKS_PER_CHECKPOINT},
+ * which is the higher ceiling we tolerate when deserializing checkpoints already accepted by L1.
+ */
+export const MAX_ATTESTABLE_BLOCKS_PER_CHECKPOINT = 72;
 
 /** Max tx effects per body (based on blob capacity per checkpoint) */
 export const MAX_TX_EFFECTS_PER_BODY = BLOBS_PER_CHECKPOINT * FIELDS_PER_BLOB;

package/src/epoch-helpers/index.ts

@@ -12,6 +12,7 @@ export type L1RollupConstants = {
   ethereumSlotDuration: number;
   proofSubmissionEpochs: number;
   targetCommitteeSize: number;
+  rollupManaLimit: number;
 };
 
 export const EmptyL1RollupConstants: L1RollupConstants = {
@@ -22,6 +23,7 @@ export const EmptyL1RollupConstants: L1RollupConstants = {
   ethereumSlotDuration: 1,
   proofSubmissionEpochs: 1,
   targetCommitteeSize: 48,
+  rollupManaLimit: Number.MAX_SAFE_INTEGER,
 };
 
 export const L1RollupConstantsSchema = zodFor<L1RollupConstants>()(
@@ -33,6 +35,7 @@ export const L1RollupConstantsSchema = zodFor<L1RollupConstants>()(
     ethereumSlotDuration: z.number(),
     proofSubmissionEpochs: z.number(),
     targetCommitteeSize: z.number(),
+    rollupManaLimit: z.number(),
   }),
 );
 
@@ -54,6 +57,25 @@ export function getSlotAtTimestamp(
     : SlotNumber.fromBigInt((ts - constants.l1GenesisTime) / BigInt(constants.slotDuration));
 }
 
+/** Returns the timestamp of the next L1 slot boundary after the given wall-clock time. */
+export function getNextL1SlotTimestamp(
+  nowInSeconds: number,
+  constants: Pick<L1RollupConstants, 'l1GenesisTime' | 'ethereumSlotDuration'>,
+): bigint {
+  const now = BigInt(nowInSeconds);
+  const elapsed = now - constants.l1GenesisTime;
+  const currentL1Slot = elapsed / BigInt(constants.ethereumSlotDuration);
+  return constants.l1GenesisTime + (currentL1Slot + 1n) * BigInt(constants.ethereumSlotDuration);
+}
+
+/** Returns the timestamp of the last L1 slot within a given L2 slot. */
+export function getLastL1SlotTimestampForL2Slot(
+  slot: SlotNumber,
+  constants: Pick<L1RollupConstants, 'l1GenesisTime' | 'slotDuration' | 'ethereumSlotDuration'>,
+): bigint {
+  return getTimestampForSlot(slot, constants) + BigInt(constants.slotDuration - constants.ethereumSlotDuration);
+}
+
 /** Returns the L2 slot number at the next L1 block based on the current timestamp. */
 export function getSlotAtNextL1Block(
   currentL1Timestamp: bigint,
@@ -140,6 +162,11 @@ export function getProofSubmissionDeadlineTimestamp(
   return getTimestampForSlot(deadlineSlot, constants);
 }
 
+/** Computes the quorum size required for a committee (⌊2n/3⌋ + 1). */
+export function computeQuorum(committeeSize: number): number {
+  return Math.floor((committeeSize * 2) / 3) + 1;
+}
+
 /** Returns the timestamp to start building a block for a given L2 slot. Computed as the start timestamp of the slot minus one L1 slot duration. */
 export function getSlotStartBuildTimestamp(
   slotNumber: SlotNumber,

package/src/file-store/factory.ts

@@ -1,9 +1,10 @@
 import { type Logger, createLogger } from '@aztec/foundation/log';
 
 import { GoogleCloudFileStore } from './gcs.js';
-import { HttpFileStore } from './http.js';
+import { HttpFileStore, type HttpFileStoreOptions } from './http.js';
 import type { FileStore, ReadOnlyFileStore } from './interface.js';
 import { LocalFileStore } from './local.js';
+import { InMemoryFileStore } from './memory.js';
 import { S3FileStore } from './s3.js';
 
 const supportedExamples = [
@@ -21,6 +22,12 @@ export async function createFileStore(
 ): Promise<FileStore | undefined> {
   if (config === undefined) {
     return undefined;
+  } else if (config.startsWith('mem://')) {
+    // In-memory store for tests (see InMemoryFileStore). The namespace is everything after the
+    // scheme, so `mem://X` and a later `mem://X` share data, like two file:// stores over one dir.
+    const namespace = config.slice('mem://'.length).replace(/\/+$/, '');
+    logger.info(`Creating in-memory file store at ${namespace}`);
+    return new InMemoryFileStore(namespace);
   } else if (config.startsWith('file://')) {
     const url = new URL(config);
     if (url.host) {
@@ -59,17 +66,26 @@ export async function createFileStore(
   }
 }
 
-export async function createReadOnlyFileStore(config: string, logger?: Logger): Promise<ReadOnlyFileStore>;
-export async function createReadOnlyFileStore(config: undefined, logger?: Logger): Promise<undefined>;
+export async function createReadOnlyFileStore(
+  config: string,
+  logger?: Logger,
+  options?: HttpFileStoreOptions,
+): Promise<ReadOnlyFileStore>;
+export async function createReadOnlyFileStore(
+  config: undefined,
+  logger?: Logger,
+  options?: HttpFileStoreOptions,
+): Promise<undefined>;
 export async function createReadOnlyFileStore(
   config: string | undefined,
   logger = createLogger('stdlib:file-store'),
+  options?: HttpFileStoreOptions,
 ): Promise<ReadOnlyFileStore | undefined> {
   if (config === undefined) {
     return undefined;
   } else if (config.startsWith('http://') || config.startsWith('https://')) {
     logger.info(`Creating read-only HTTP file store at ${config}`);
-    return new HttpFileStore(config, logger);
+    return new HttpFileStore(config, logger, options);
   } else {
     return await createFileStore(config, logger);
   }

package/src/file-store/http.ts

@@ -10,6 +10,14 @@ import { pipeline } from 'stream/promises';
 
 import type { ReadOnlyFileStore } from './interface.js';
 
+/** Options for configuring HttpFileStore behavior. */
+export interface HttpFileStoreOptions {
+  /** Retry backoff intervals in seconds. Empty array disables retries. Default: [1, 1, 3]. */
+  retryBackoff?: number[];
+  /** Request timeout in milliseconds. Default: no timeout (axios default). */
+  timeoutMs?: number;
+}
+
 export class HttpFileStore implements ReadOnlyFileStore {
   private readonly axiosInstance: AxiosInstance;
   private readonly fetch: <T>(config: AxiosRequestConfig) => Promise<AxiosResponse<T>>;
@@ -17,17 +25,28 @@ export class HttpFileStore implements ReadOnlyFileStore {
   constructor(
     private readonly baseUrl: string,
     private readonly log: Logger = createLogger('stdlib:http-file-store'),
+    options?: HttpFileStoreOptions,
   ) {
-    this.axiosInstance = axios.create();
-    this.fetch = async <T>(config: AxiosRequestConfig) => {
-      return await retry(
-        () => this.axiosInstance.request<T>(config),
-        `Fetching ${config.url}`,
-        makeBackoff([1, 1, 3]),
-        this.log,
-        /*failSilently=*/ true,
-      );
-    };
+    this.axiosInstance = axios.create({
+      ...(options?.timeoutMs !== undefined && { timeout: options.timeoutMs }),
+    });
+
+    const retryBackoff = options?.retryBackoff ?? [1, 1, 3];
+    if (retryBackoff.length > 0) {
+      this.fetch = async <T>(config: AxiosRequestConfig) => {
+        return await retry(
+          () => this.axiosInstance.request<T>(config),
+          `Fetching ${config.url}`,
+          makeBackoff(retryBackoff),
+          this.log,
+          /*failSilently=*/ true,
+        );
+      };
+    } else {
+      this.fetch = async <T>(config: AxiosRequestConfig) => {
+        return await this.axiosInstance.request<T>(config);
+      };
+    }
   }
 
   public async read(pathOrUrl: string): Promise<Buffer> {

package/src/file-store/index.ts

@@ -1,2 +1,4 @@
 export * from './interface.js';
 export * from './factory.js';
+export { type HttpFileStoreOptions } from './http.js';
+export { InMemoryFileStore } from './memory.js';

package/src/file-store/memory.ts

@@ -0,0 +1,94 @@
+import { writeFile } from 'fs/promises';
+import { promisify } from 'util';
+import { gunzip as gunzipCb, gzip as gzipCb } from 'zlib';
+
+import type { FileStore, FileStoreSaveOptions } from './interface.js';
+
+const gzip = promisify(gzipCb);
+const gunzip = promisify(gunzipCb);
+
+// Backing data is shared per namespace so that a store created for writing and a (read-only) store
+// created later from the same `mem://` URL observe the same files — mirroring how two `file://`
+// stores over the same directory share state on disk. Without this, `FileStoreTxSource`, which
+// builds its own store from the URL, would never see what a test wrote through a different instance.
+const namespaces = new Map<string, Map<string, Buffer>>();
+
+/**
+ * In-memory {@link FileStore}, addressed via `mem://<namespace>/...`. Reads and writes are synchronous
+ * map operations with no disk I/O, so it is deterministic — useful for tests that exercise consumers
+ * of a file store (e.g. the tx file store) rather than the on-disk store itself, where a real
+ * filesystem can introduce timing flakes.
+ */
+export class InMemoryFileStore implements FileStore {
+  private readonly files: Map<string, Buffer>;
+
+  constructor(private readonly namespace: string) {
+    let files = namespaces.get(namespace);
+    if (!files) {
+      files = new Map();
+      namespaces.set(namespace, files);
+    }
+    this.files = files;
+  }
+
+  /** Clears all in-memory file store data, or a single namespace. Intended for test isolation. */
+  static clear(namespace?: string): void {
+    if (namespace === undefined) {
+      namespaces.clear();
+    } else {
+      namespaces.delete(namespace);
+    }
+  }
+
+  public async save(path: string, data: Buffer, opts?: FileStoreSaveOptions): Promise<string> {
+    const toStore = opts?.compress ? await gzip(data) : data;
+    const key = this.key(path);
+    this.files.set(key, Buffer.from(toStore));
+    return `mem://${this.namespace}/${key}`;
+  }
+
+  public async upload(destPath: string, srcPath: string, opts?: FileStoreSaveOptions): Promise<string> {
+    const { readFile } = await import('fs/promises');
+    return this.save(destPath, await readFile(srcPath), opts);
+  }
+
+  public async read(pathOrUrl: string): Promise<Buffer> {
+    const data = this.files.get(this.key(pathOrUrl));
+    if (data === undefined) {
+      throw new Error(`File not found in memory store: ${pathOrUrl}`);
+    }
+    // Match LocalFileStore: transparently gunzip content that was stored compressed.
+    if (data.length >= 2 && data[0] === 0x1f && data[1] === 0x8b) {
+      return await gunzip(data);
+    }
+    return data;
+  }
+
+  public async download(pathOrUrl: string, destPath: string): Promise<void> {
+    await writeFile(destPath, await this.read(pathOrUrl));
+  }
+
+  public exists(pathOrUrl: string): Promise<boolean> {
+    return Promise.resolve(this.files.has(this.key(pathOrUrl)));
+  }
+
+  /** Lists stored file keys, optionally restricted to those under `prefix`. Not part of {@link FileStore}. */
+  public listFiles(prefix = ''): string[] {
+    const keys = [...this.files.keys()];
+    const normalizedPrefix = prefix.replace(/^\/+/, '');
+    return normalizedPrefix ? keys.filter(k => k.startsWith(normalizedPrefix)) : keys;
+  }
+
+  /** Resolves a relative path or a `mem://` URI (as returned by `save`) to the registry key. */
+  private key(pathOrUrl: string): string {
+    const ownPrefix = `mem://${this.namespace}/`;
+    if (pathOrUrl.startsWith(ownPrefix)) {
+      return pathOrUrl.slice(ownPrefix.length);
+    }
+    if (pathOrUrl.startsWith('mem://')) {
+      // A mem:// URI for some namespace: drop the scheme and the leading namespace segment.
+      return pathOrUrl.slice('mem://'.length).replace(/^[^/]*\//, '');
+    }
+    return pathOrUrl.replace(/^\/+/, '');
+  }
+}

package/src/gas/README.md

@@ -0,0 +1,318 @@
+# Aztec Gas and Fee Model
+
+The minimum fee per mana and its components are computed on L1 in
+`l1-contracts/src/core/libraries/rollup/FeeLib.sol` (`fee_math.ts` in this directory is a
+TypeScript port of those formulas, used to predict fees a few slots ahead). This document
+describes the formulas, the oracle lag/lifetime mechanism, how a transaction's fee is
+derived from them, the gas and data limits, and the TypeScript types in this directory.
+
+## Mana
+
+Aztec meters work as gas in two dimensions: **DA** (data availability, i.e. blob data
+published to L1) and **L2** (execution). The L2 dimension is called **mana** (analogous to
+Ethereum gas): block headers track `totalManaUsed`, and the fee model below prices one unit
+of mana. The total fee is `gasUsed * feePerGas` summed across both dimensions.
+
+### The DA dimension is priced at zero
+
+Only the L2 dimension currently carries a price. When building checkpoint global variables,
+the sequencer sets `feePerDaGas = 0` and sets `feePerL2Gas` to the L1-computed minimum fee
+per mana (`sequencer-client/src/global_variable_builder/global_builder.ts` and
+`fee_provider.ts` in the same directory). The cost of publishing data is still recovered:
+the blob gas a checkpoint pays on L1 is part of the *sequencer cost* component of the mana
+fee below. DA gas remains fully metered and limited (see
+[Gas and Data Limits](#gas-and-data-limits)) — it bounds how much data a tx or checkpoint
+may publish — it just contributes nothing to the fee today.
+
+## Fee Components
+
+The minimum fee per mana has four components. All are computed in ETH (wei) per mana and
+converted to the fee asset at the end (see [Fee Asset Price](#fee-asset-price)).
+
+### Sequencer Cost
+
+L1 cost to propose a checkpoint (calldata gas + blob data), amortized over `manaTarget`:
+
+```
+sequencerCost = ceil(((L1_GAS_PER_CHECKPOINT_PROPOSED * baseFee)
+              + (BLOBS_PER_CHECKPOINT * BLOB_GAS_PER_BLOB * blobFee))
+              / manaTarget)
+```
+
+Note that `BLOBS_PER_CHECKPOINT` here is FeeLib's own constant (3), not the protocol blob
+capacity of the same name (6) — see the note under [Key Constants](#key-constants).
+
+### Prover Cost
+
+L1 cost to verify an epoch proof, amortized over epoch duration and `manaTarget`, plus a
+governance-set proving cost that compensates for off-chain proof generation:
+
+```
+proverCost = ceil(ceil((L1_GAS_PER_EPOCH_VERIFIED * baseFee) / epochDuration) / manaTarget)
+           + provingCostPerMana
+```
+
+Updates to `provingCostPerMana` are rate-limited on L1 (`FeeLib.updateProvingCostPerMana`):
+at most one update every 30 days, each moving the value by at most ×1.5 (or ÷1.5), with a
+floor of 2 wei per mana.
+
+### Congestion Cost
+
+An exponential surcharge when the network is congested (inspired by EIP-1559; the
+implementation uses the `fakeExponential` Taylor series approximation from EIP-4844):
+
+```
+baseCost       = sequencerCost + proverCost
+congestionCost = floor(baseCost * congestionMultiplier / MINIMUM_CONGESTION_MULTIPLIER) - baseCost
+```
+
+When there is no congestion the multiplier equals `MINIMUM_CONGESTION_MULTIPLIER` (1e9)
+and congestion cost is zero.
+
+### Congestion Multiplier
+
+```
+excessMana           = max(0, prevExcessMana + prevManaUsed - manaTarget)
+denominator          = manaTarget * 854,700,854 / 1e8    ≈ 8.547 * manaTarget
+congestionMultiplier = fakeExponential(MINIMUM_CONGESTION_MULTIPLIER,
+                                       min(excessMana, 100 * denominator), denominator)
+```
+
+Each additional `manaTarget` of excess mana multiplies the fee by `e^(1/8.547) ≈ 1.124`,
+i.e. ~12.5%. The exponent is capped at 100 (multiplier ≤ ~2.7e43 × the minimum) to keep
+the Taylor series from overflowing.
+
+### Total
+
+```
+minFeePerMana = sequencerCost + proverCost + congestionCost
+```
+
+Each component is converted from ETH to the fee asset individually (rounding up) before
+summing. The sum is capped at `type(uint128).max` (`FeeLib.summedMinFee`) so it always fits
+the proposal header's `feePerL2Gas` field — without the cap, extreme congestion could
+produce a fee no valid header can represent, halting the chain.
+
+## L1 Gas Oracle: Lag and Lifetime
+
+The oracle feeds Ethereum's `baseFee` and `blobFee` into the fee model using a two-phase
+(`pre` / `post`) system that smooths out L1 fee volatility.
+
+- **LAG = 2 slots** — when new L1 fees are observed, they activate `LAG` slots later
+  (`slotOfChange = currentSlot + LAG`). This gives mempool transactions time to land
+  before fees change.
+- **LIFETIME = 5 slots** — after an oracle update, further updates are ignored
+  (`updateL1GasFeeOracle` returns without effect) until `slotOfChange + (LIFETIME - LAG)`
+  = 3 more slots have passed. This rate-limits how frequently L1 fee data can change.
+
+Fee resolution at a given timestamp:
+
+```
+if slot < slotOfChange  →  use pre  (old fees)
+else                    →  use post (new fees)
+```
+
+**Net effect**: L1 fee changes reach L2 with a 2-slot delay and can update at most once
+every 5 slots.
+
+Because queued values only activate `LAG` slots later, the min fee for the next `LAG`
+slots is fully determined by current on-chain state. `fee_math.ts` ports the FeeLib
+formulas so the sequencer can predict min fees over that window (`FeePredictor` in
+`sequencer-client/src/global_variable_builder`), under a configurable mana-usage
+assumption (`ManaUsageEstimate`: none / target / limit).
+
+### Worked Example
+
+Suppose the oracle is updated at slot 10 with new L1 fees. Here is the timeline:
+
+```
+Slot  Oracle state          Active fees   Notes
+────  ────────────────────  ────────────  ──────────────────────────────────
+ 10   pre=A, post=B, soc=12   A           Update queued. slotOfChange = 10 + LAG = 12.
+ 11   (same)                  A           Still before slotOfChange → pre (A).
+ 12   (same)                  B           slot >= slotOfChange → post (B) activates.
+ 13   (same)                  B           B remains active.
+ 14   (same)                  B           B remains active.
+ 15   Update allowed again    B           Earliest next update: soc + (LIFETIME - LAG)
+                                          = 12 + 3 = 15.
+```
+ 
+Key observations:
+
+1. **Slots 10-11**: The old fees (A) are still in effect. Transactions submitted during
+   these slots see the old L1 cost. This is the **LAG** window — it gives pending
+   transactions 2 slots to land before fees change.
+
+2. **Slot 12**: The new fees (B) activate. Any checkpoint proposed at slot >= 12 uses B
+   for its sequencer/prover cost calculation.
+
+3. **Slots 12-14**: No new oracle update is accepted. The system is in a **cooldown**
+   period of `LIFETIME - LAG = 3` slots after the transition.
+
+4. **Slot 15**: A new oracle update can be queued (earliest `acceptableSlot`). If
+   triggered, the new values would activate at slot 15 + LAG = 17.
+
+## Fee Asset Price
+
+Fees are computed in ETH (wei) internally and converted to the fee asset (Fee Juice) via
+`ethPerFeeAsset` (1e12 precision), rounding up (`PriceLib.toFeeAsset` in
+`l1-contracts/src/core/libraries/compressed-data/fees/FeeConfig.sol`).
+
+Each checkpoint proposal carries a fee-asset price modifier (`OracleInput`) chosen by the
+proposer, bounded to ±1% (±100 bps) per checkpoint:
+
+```
+newPrice = currentPrice * (10000 + modifierBps) / 10000
+```
+
+The result is clamped to [`MIN_ETH_PER_FEE_ASSET` = 100, `MAX_ETH_PER_FEE_ASSET` = 1e14],
+i.e. 1e-10 to 100 ETH per fee asset. The floor of 100 guarantees a ±1% step always moves
+the integer price by at least 1.
+
+## Transaction Fees
+
+The checkpoint's `gasFees` (the L1-computed min fee per mana, with DA priced at zero) act
+as a base fee. Senders declare `GasSettings`: gas limits, teardown gas limits,
+`maxFeesPerGas`, and `maxPriorityFeesPerGas`. A tx is only includable if `maxFeesPerGas`
+covers the checkpoint's `gasFees` in both dimensions. The effective fee adds an
+EIP-1559-style priority fee on top of the base, capped by the max
+(`computeEffectiveGasFees` in `stdlib/src/fees/transaction_fee.ts`):
+
+```
+effectiveFeePerGas = gasFees + min(maxPriorityFeePerGas, maxFeesPerGas - gasFees)   (per dimension)
+transactionFee     = billedGas.daGas * effectiveFeePerDaGas + billedGas.l2Gas * effectiveFeePerL2Gas
+```
+
+`billedGas` is actual consumption except for the teardown phase, which is billed at the
+declared `teardownGasLimits` rather than actual usage — the fee is charged during teardown
+itself, before actual teardown consumption is known.
+
+## Maximum Fee Change Rate
+
+| Component              | Bound                                                    |
+| ---------------------- | -------------------------------------------------------- |
+| L1 base fee / blob fee | At most once every 5 slots (oracle LIFETIME)             |
+| Fee asset price        | ±1% per checkpoint                                       |
+| Proving cost per mana  | At most ×1.5 (or ÷1.5) per update, one update per 30 days |
+| Congestion multiplier  | Depends on excess mana accumulation/drain per checkpoint |
+| Sequencer/prover costs | Scale linearly with L1 fees                              |
+
+## Key Constants
+
+| Constant                        | Value          |
+| ------------------------------- | -------------- |
+| `L1_GAS_PER_CHECKPOINT_PROPOSED` | 300,000       |
+| `L1_GAS_PER_EPOCH_VERIFIED`     | 3,600,000      |
+| `BLOBS_PER_CHECKPOINT` (FeeLib) | 3              |
+| `BLOB_GAS_PER_BLOB`             | 2^17           |
+| `MINIMUM_CONGESTION_MULTIPLIER` | 1e9            |
+| `LAG`                           | 2 slots        |
+| `LIFETIME`                      | 5 slots        |
+| `MIN_ETH_PER_FEE_ASSET`         | 100 (1e-10 ETH) |
+| `MAX_ETH_PER_FEE_ASSET`         | 1e14 (100 ETH) |
+
+⚠️ Name clash: `FeeLib.sol` (and its port `fee_math.ts`) defines `BLOBS_PER_CHECKPOINT = 3`,
+used only to price the sequencer's blob costs. The protocol constant of the same name in
+`@aztec/constants` is **6** — the actual blob capacity of a checkpoint, used throughout the
+limits section below. The FeeLib value is a holdover from the pre-checkpoint
+`BLOBS_PER_BLOCK`, so the fee model prices half of a full checkpoint's blobs.
+
+## Gas and Data Limits
+
+The fee model above is *how much you pay* per unit of gas; this section is *how much you may use*. Limits
+form a hierarchy from a single transaction up to a whole checkpoint, and a tx that is admissible for relay
+must also be buildable into a block and fit a valid checkpoint.
+
+### Per-tx protocol maxima
+
+Hard ceilings on what any single tx may declare, independent of network configuration. Declaring more is
+rejected everywhere a tx is validated.
+
+- **`MAX_TX_DA_GAS`** (271,200) — `MAX_TX_BLOB_DATA_SIZE_IN_FIELDS` (8,475) × `DA_GAS_PER_FIELD` (32). This
+  is the most DA a single tx's effects can encode into a blob, so it is the most DA gas a tx could ever use.
+  Defined in `constants/src/constants.ts`.
+- **`MAX_PROCESSABLE_L2_GAS`** (6,540,000) — the AVM's maximum processable L2 gas, derived in Noir as
+  `PUBLIC_TX_L2_GAS_OVERHEAD + AVM_MAX_PROCESSABLE_L2_GAS` (`constants/src/constants.gen.ts`).
+
+### Network admission limits
+
+The most a single tx may *declare* and still be relayed across the network. Computed by
+`computeNetworkTxGasLimits` in `tx_gas_limits.ts` per dimension as:
+
+```
+min(per-tx max, ceil(checkpointBudget / blocksPerCheckpoint * minMultiplier))
+```
+
+The per-block share mirrors what a proposer grants the first block of a checkpoint
+(`CheckpointBuilder.capLimitsByCheckpointBudgets`), so a tx declaring this much is packable into a block.
+The network-minimum multipliers are `MIN_PER_BLOCK_ALLOCATION_MULTIPLIER` (1.2, L2 and tx count) and
+`MIN_PER_BLOCK_DA_ALLOCATION_MULTIPLIER` (1.5, DA). DA's is higher so a maximal contract class
+registration (~97k DA gas) fits a single block at mainnet geometry (72s slots, 6s blocks → 10 blocks per
+checkpoint).
+
+The DA budget is `getDaCheckpointBudgetForTxs(maxBlocksPerCheckpoint)`, not the raw
+`MAX_PROCESSABLE_DA_GAS_PER_CHECKPOINT` (786,432). Blob encoding spends overhead fields that no tx pays DA
+gas for — one checkpoint-end marker field and the per-block block-end fields (7 for the first block, 6 for
+each subsequent block, `blob-lib/src/encoding/block_blob_data.ts`) — so the raw constant is unattainable. The
+getter nets out the full overhead for a checkpoint of `maxBlocksPerCheckpoint` blocks: at mainnet geometry
+(10 blocks) that is `(24,576 − 1 − 7 − 9×6) × 32 = 24,514 × 32 = 784,448` DA gas. Subtracting every block's
+overhead (not just the first) keeps admission at or below the builder's first-block blob-field cap at every
+geometry — the builder is the most generous for the first block (it only reserves that block's own block-end
+overhead), so being conservative here is what guarantees admitted ⇒ buildable. Without this netting a tx
+near the raw limit would be admitted but never buildable.
+
+These limits depend on network-wide inputs only (timetable-derived blocks-per-checkpoint, checkpoint
+budgets, the network-minimum multipliers), never on a node's local restrictiveness. Every node always
+advertises them in `NodeInfo.txsLimits` (a required field); wallets read it and pass `txsLimits.gas` to
+`GasSettings.fallback` as the default gas limits when sending without explicit limits, and they are enforced
+by `GasLimitsValidator` (clamped to the per-tx protocol maxima) at three points: RPC tx acceptance
+(`aztec-node/src/aztec-node/server.ts`), gossip validation (`p2p/src/services/libp2p/libp2p_service.ts`),
+and pending-pool admission (`p2p/src/client/factory.ts`). They are deliberately *not* enforced at req/resp or
+block-proposal validation — admission is relay policy, not block validity.
+
+### Per-block builder budgets
+
+While packing a checkpoint, `CheckpointBuilder.capLimitsByCheckpointBudgets`
+(`validator-client/src/checkpoint_builder.ts`) computes each block's budget as a fair share of the remaining
+checkpoint budget across the remaining blocks, scaled by the configured multipliers. Operators may raise the
+multipliers above the network minimums but not lower them — the sequencer fails startup otherwise
+(`assertConfigMeetsNetworkTxLimits` in `sequencer-client/src/sequencer/sequencer.ts`), since a node that
+allocates less than it admits would accept txs over RPC/gossip that its builder can never pack.
+
+The fair share is then min'ed with the operator's absolute per-block caps `maxL2BlockGas` / `maxDABlockGas`
+and the blob-field cap (checkpoint capacity net of the checkpoint-end marker and this block's block-end
+overhead). The absolute caps are allowed to be restrictive: a cap below the network admission limit only
+produces a startup warning, not a failure, because such txs simply stay in the pool for other proposers to
+include.
+
+### Per-checkpoint budgets
+
+The outermost limits, enforced as proposal validity in `validateCheckpointLimits`
+(`stdlib/src/checkpoint/validate.ts`) and physically by blob encoding:
+
+- **Mana** — total L2 gas across all blocks ≤ `rollupManaLimit` (= `manaTarget × 2` on L1,
+  `l1-contracts/src/core/libraries/rollup/FeeLib.sol`).
+- **DA gas** — total DA gas ≤ raw `MAX_PROCESSABLE_DA_GAS_PER_CHECKPOINT` (786,432).
+- **Blob fields** — total ≤ `BLOBS_PER_CHECKPOINT × FIELDS_PER_BLOB` (6 × 4,096 = 24,576).
+- **Tx counts** — total txs ≤ `maxTxsPerCheckpoint` when configured.
+
+### Summary
+
+| Limit                                   | Value (mainnet defaults)        | Scope         | Where enforced                                              |
+| --------------------------------------- | ------------------------------- | ------------- | ----------------------------------------------------------- |
+| `MAX_TX_DA_GAS`                         | 271,200                         | per-tx        | every gas validator (hard ceiling)                          |
+| `MAX_PROCESSABLE_L2_GAS`                | 6,540,000                       | per-tx        | every gas validator (hard ceiling)                          |
+| Network DA admission limit              | min(271,200, ceil(784,448/10×1.5)) = 117,668 | per-tx (relay) | RPC, gossip, pending pool (`GasLimitsValidator`)         |
+| Network L2 admission limit              | min(6,540,000, ceil(manaLimit/10×1.2)) | per-tx (relay) | RPC, gossip, pending pool (`GasLimitsValidator`)         |
+| Per-block fair share + caps             | remaining budget / blocks × multiplier, min absolute caps & blob-field cap | per-block | `CheckpointBuilder.capLimitsByCheckpointBudgets`         |
+| `rollupManaLimit`                       | `manaTarget × 2`                | per-checkpoint | `validateCheckpointLimits`                                |
+| `MAX_PROCESSABLE_DA_GAS_PER_CHECKPOINT` | 786,432                         | per-checkpoint | `validateCheckpointLimits` + blob encoding                |
+| `BLOBS_PER_CHECKPOINT × FIELDS_PER_BLOB`| 24,576                          | per-checkpoint | `validateCheckpointLimits` + blob encoding                |
+
+## TypeScript Types
+
+- **`Gas`** — gas quantity in two dimensions (`daGas`, `l2Gas`).
+- **`GasFees`** — per-unit price in each dimension (`feePerDaGas`, `feePerL2Gas`).
+- **`GasSettings`** — sender-chosen fee parameters: gas limits, teardown limits, max fees, priority fees.
+- **`GasUsed`** — actual consumption after execution. Note: `billedGas` uses the teardown gas *limit*, not actual usage.
+- **`fee_math.ts`** — TypeScript port of the FeeLib formulas, used for fee prediction over the oracle LAG window.