fix(e2e): ground-truth pool verdict + earned Tari lag tolerance (#746)

[VijitSingh97]

Closes #746. The two Tor-timing flakes that hit 2 of 4 e2e runs during the v1.12.0 cut (neither a code defect — the same ref passed 374/0 and 380/0 on clean runs).

Pool type — check the ground truth, not just the classifier

After a sidechain switch, p2pool runs the NEW --mini/--nano flags while the dashboard's peer-port classifier can still report the pre-switch pool — determinate, wrong, and transient. Both assert_scenario and assert_pool_switched hard-failed that as a "render bug". New shared assert_pool_type adds the check the harness never made: the rendered P2POOL_FLAGS.

got vs want	P2POOL_FLAGS	verdict
match	—	pass
Unknown/empty	—	warn (peer timing, #454)
wrong	correct	warn (classifier lag, #746)
wrong	wrong	fail (real render bug)

The mismatch path now verifies actual flags, so this is a stronger check than the old hard-fail.

Tari — earned lag tolerance

Every per-scenario restart sends Tari back through "discovering the target height" (loading), which can outlast wait_tari_synced over Tor; the bare assert_eq then fired on an in-progress state (the cold-read anti-pattern the harness's own #687 comment describes — it passed in 3 of 4 scenarios of the failing run). New assert_tari_synced_required: done passes and records TARI_SEEN_DONE; loading/syncing after that proof warns as re-discovery lag; a Tari that never syncs still fails the gate — the tolerance is earned, not free.

Validation

• Harness selftest: 148/0, covering every new branch (flags ground truth per pool, all four pool verdicts, the earned-tolerance matrix incl. first-look-loading → FAIL, and the TARI_SEEN_DONE side effect).
• make lint-sh clean.
• Live targeted e2e with this branch (gouda + borrowed miner-1) running now; result will be posted here before merge.

[VijitSingh97]

Live validation complete: targeted e2e with this branch on gouda + borrowed miner-1 — 380 passed / 0 failed / 3 skipped, ✓ E2E PASSED. The shared assert_pool_type warn path fired once live (Unknown → peer-timing WARN, #454), exercising the refactored verdict on the real bench; the classifier-lag and Tari-tolerance branches are covered by the selftest (148/0). Clean restore (miner repointed, baseline healthy).