tests/shadow-testing: adopt new relayer flags in follow mode

- relayer.json.template: disable_l2_watcher + chain_nonce_only
- 10-follow-up.sh --reset: TRUNCATE pending_transaction (stale rows
  poison sender nonces and can replay old calldata on a fresh fork)
- mark Traps 26/7/9 as fixed upstream; document the new config knobs
  in follow/GUIDE.md

Author: lispc

tests/shadow-testing/docs/COMMON-TROUBLESHOOTING.md

@@ -73,6 +73,9 @@ Before executing a single command:
 - **Rule**: Use **internal** L2 RPC proxies only.
 
 ### Trap 7: Relayer Nonce Desync
+
+> **Fixed upstream / harness**: the relayer sender config now supports `chain_nonce_only` (default false), which initializes the nonce from the chain pending nonce only, ignoring `pending_transaction` — the shadow relayer config template sets it. Additionally `10-follow-up.sh --reset` now TRUNCATEs `pending_transaction` (the rows are not chain/fork-scoped and can also replay old calldata onto a fresh fork). The manual fix below is only needed if you run a relayer without those.
+
 - **Symptom**: Tx sent but never mined; `eth_getTransactionReceipt` returns null forever.
 - **Cause**: `pending_transaction` table retains nonces from previous runs that were never confirmed. Relayer initializes nonce from `maxDbNonce + 1`, which is ahead of the on-chain nonce.
 - **Rule**: After any relayer crash or Anvil restart:
@@ -82,6 +85,9 @@ Before executing a single command:
   Then restart the relayer.
 
 ### Trap 9: Anvil `eth_estimateGas` Rejects Fee Caps
+
+> **Fixed upstream**: `rollup/internal/controller/sender/estimategas.go` now sends an explicit non-zero gas cap in the estimation `CallMsg`, so Anvil no longer rejects fee-capped estimate calls. This entry is kept for reference when running older relayer builds.
+
 - **Symptom**: `failed to get fee data, err: Out of gas: gas required exceeds allowance: 0`.
 - **Cause**: Anvil's `eth_estimateGas` fails when `CallMsg` has `GasFeeCap`/`GasTipCap` set but `Gas` is 0 (Go Ethereum client's default).
 - **Rule**: If you hit this on a shadow fork, the correct fix belongs in the upstream `rollup/internal/controller/sender/estimategas.go` (do not maintain a local patch in this branch). Verify with the latest `develop` code and, if still present, fix it there so all shadow tests benefit.

tests/shadow-testing/follow/GUIDE.md

@@ -68,7 +68,7 @@ One command brings up the entire follow-mode stack (`scripts/10-follow-up.sh`):
 3. **Verifier wrapper** — deploy/register the `ZkEvmVerifierPostFeynman` wrapper matching the guest under test (manual procedure: "Real Verifier Deployment" in the [Snapshot Replay Mode Guide](../snapshot/GUIDE.md#real-verifier-deployment)).
 4. **Coordinator** — start `coordinator_api` + `coordinator_cron` against the shadow DB.
 5. **Provers** — start one prover per GPU.
-6. **Relayer** — start the rollup relayer with local proposers **disabled** (`l2_config.{chunk,batch,bundle}_proposer_config.disable = true`): the relayer then only commits/finalizes what the DB contains, instead of synthesizing bundles at an unrealistic rate.
+6. **Relayer** — start the rollup relayer with local proposers **disabled** (`l2_config.{chunk,batch,bundle}_proposer_config.disable = true`): the relayer then only commits/finalizes what the DB contains, instead of synthesizing bundles at an unrealistic rate. The shadow relayer config template (`lib/configs/relayer.json.template`) additionally sets `l2_config.disable_l2_watcher = true` (the poll sync, not the relayer's L2 watcher, keeps `l2_block` populated — Trap 26) and `sender_config.chain_nonce_only = true` (sender nonces initialize from the fork, never from stale `pending_transaction` rows — Trap 7/27).
 7. **Background daemons** — poll-sync, sweeper, hourly monitor (plain background processes with pidfiles in `.work/`; no agent/session cron is required).
 
 The run window is recorded in `.work/follow-run.env` (`SHADOW_REPORT_START`, `FOLLOW_RUN_HOURS=48`). Tear down with `make follow-stop` (`scripts/11-follow-stop.sh`, supports `--keep-anvil`).

tests/shadow-testing/follow/TROUBLESHOOTING.md

@@ -78,6 +78,8 @@ see also the [Follow Mode Guide](./GUIDE.md). Trap numbers are stable across the
 
 ### Trap 26: Relayer L2 Watcher Genesis-Crawls an Empty `l2_block` [follow mode]
 
+> **Fixed upstream**: the relayer now supports `l2_config.disable_l2_watcher` (default false = production behavior unchanged), which skips the watcher loop entirely. The shadow relayer config template (`lib/configs/relayer.json.template`) sets it — in follow mode the poll sync is what keeps `l2_block` populated for newly synced chunks, so the watcher is redundant. The harness ordering below remains as defense-in-depth.
+
 - **Symptom**: `l2_block` fills with rows numbered 1, 2, 3, … at ~30–60 blocks/s, `chunk_hash = NULL`, `withdraw_root = 0x0000…` — a full L2 history crawl that would take days and tens of GB, while relayer logs show `retrieved block height=…` climbing from genesis.
 - **Cause**: `rollup_relayer` always runs an L2 watcher loop (`rollup/cmd/rollup_relayer/app/app.go` → `TryFetchRunningMissingBlocks` every 2 s) that fetches **all** blocks from `MAX(l2_block.number)+1` up to the L2 tip — reading the max **once** at loop entry. If the relayer is alive while `--reset` TRUNCATEs `l2_block` (or the table is empty for any other reason), the watcher sees `MAX = 0` and starts crawling from block 1; the in-flight call keeps going for days even after the baseline repopulates the window.
 - **Fix (in place)**: `10-follow-up.sh --reset` now runs `11-follow-stop.sh` **before** the TRUNCATE, so no writer survives into the reset. Never TRUNCATE `l2_block` (or let it go empty) while a relayer is running.

tests/shadow-testing/follow/scripts/10-follow-up.sh

@@ -178,10 +178,16 @@ if $RESET_DB; then
     if ! $DRY_RUN; then
         bash "${SCRIPT_DIR}/11-follow-stop.sh" || log_warn "  11-follow-stop returned non-zero (continuing)"
     fi
-    log_warn "  --reset: truncating chunk/batch/bundle/l2_block/l1_message"
+    log_warn "  --reset: truncating chunk/batch/bundle/l2_block/l1_message/pending_transaction"
     if ! $DRY_RUN; then
         psql "$SHADOW_DSN" -Atq --set ON_ERROR_STOP=1 -c "TRUNCATE chunk, batch, bundle, l2_block, l1_message" >/dev/null \
             || { log_error "TRUNCATE failed — aborting before baseline"; exit 1; }
+        # Trap 7/27: pending_transaction rows are not chain/fork-scoped. Stale
+        # rows from a previous run poison the relayer sender's nonce
+        # (initializeNonce = max(db_max+1, chain_nonce)) and can replay old
+        # calldata onto the fresh fork. Wipe them on every reset.
+        psql "$SHADOW_DSN" -Atq --set ON_ERROR_STOP=1 -c "TRUNCATE pending_transaction" >/dev/null \
+            || { log_error "TRUNCATE pending_transaction failed — aborting"; exit 1; }
         # Trap 21: wipe prover-local proof caches; they hold proofs from the
         # previous run/circuit version and get replayed as VData mismatches.
         rm -rf "${WORK_DIR}"/prover-*/db

tests/shadow-testing/lib/configs/relayer.json.template

@@ -1,6 +1,7 @@
 {
   "l2_config": {
     "endpoint": "{{L2_ENDPOINT}}",
+    "disable_l2_watcher": true,
     "relayer_config": {
       "sender_config": {
         "endpoint": "{{ANVIL_RPC}}",
@@ -16,7 +17,8 @@
         "min_gas_tip": 0,
         "check_pending_time": 60,
         "max_pending_blob_txs": 3,
-        "fusaka_timestamp": 9999999999
+        "fusaka_timestamp": 9999999999,
+        "chain_nonce_only": true
       },
       "commit_sender_signer_config": {
         "signer_type": "PrivateKey",