[futurepaul]

Flake: hermes' "and each is charged on its owner's ledger" failed with []. Run alone on master (cargo xtask e2e --only hermes, 5ea1a09) it failed again here: 68 passed, 1 failed.

Cause (test): a model call is charged after its answer. The charge goes through the fragment-ledger queue (cell/src/meter.rs charge), whose consumer batches for up to 1 s (max_batch_timeout: 1). The check read the owner's ledger once, right after the turn. The node log shows the check's POST /api/test/ledger before the queue's first QUEUE fragment-ledger 3/3 delivery, so no charge had landed yet. When the whole lane runs, other sections leave time for the batch, which is why it passed there. The vision check later in the section reads the same way (vision_entries).

Fix: both checks wait for the condition itself (s.eventually, the section's TURN bound). The first waits until the owner's ledger has a charged row for each model call made (aig.len() >= calls.len()), which is stricter than before ("not empty"). The second waits until each vision call has its charged row. They wait on the rows, never a fixed time.

Proof: cargo xtask e2e --only hermes: 69 passed, 0 failed (master: 68/1).

🤖 Generated with Claude Code

[alexlwn123]

Pass at 9783d57f4a4d1e7da08d440eb7d04f9b98cfa435.

Both ledger checks now wait for asynchronous charge delivery using the existing bounded polling helper. The first also requires enough charged rows for the observed calls; the vision check retains its agent/model filter and nonempty-call assertion. This is a small, appropriate test-only fix, with no blocking findings.

Validation: cargo check -p fragment-e2e --locked and git diff HEAD^ HEAD --check passed. Traced the polling helper, ledger readers, and charge queue configuration. CI is still pending; I did not rerun the full Docker/Hermes lane (the PR reports 69/69 passing). The referenced external engineering-style file is unavailable on this machine.