cloud-lab: combined pytest run red on main — cumulative in-memory sessions break fee/allotment assertions; clientd offer-string drift #792

The standard combined suite invocation on main (cb4db7f, verified 2026-10-09 on the ai-legion lab host; identical shape on PR #771 head b3ce6d5) fails 6 tests:

• test_fee_is_sum_then_ceil[255/1023/2047], test_above_swap_fee_succeeds_with_fee_deducted, test_fee_mint_payments_do_not_break_the_free_mint_path: the session event's allotment is cumulative per client MAC (in-memory sessions — the cloud-lab README already notes "sessions are in-memory per client MAC"), so per-payment assertions break once earlier payments or the clientd renewal daemon (test_clientd_autotopup.py runs earlier alphabetically in the same invocation) have accumulated time on the static client MAC 02:00:00:00:00:20. Attribution verified empirically: [255] passes on a freshly restarted upstream, and a main-built upstream fails identically to the PR head — pre-existing on main, not a fix(merchant): boot resume replays journaled swap intents — close the #719 crash window (#497) #771 regression.
• test_list_offers_parses_real_advertisement: expects 1 sats/step, scripts/tollgate-clientd.py prints 1 sat/step.

Impact: the combined client pytest -sv invocation cannot serve as a green gate — the next lane that runs it as-is will re-triage the same six failures.

Suggested directions: per-file randomized client MACs (precedent: _mac() in test_external_mints.py), a session-reset fixture between files, or documenting a canonical segmented invocation; align the clientd string expectation.

Evidence (lane host, session rc1-771): /root/src/rel-771-logs/std-suite-pytest.log, attribution runs described in /root/src/rel-771-transcript.md §4.5, summary in /root/src/rel-771-handoff.md Appendix B.