Complete an unattended Knight level 8-20 progression soak

[adrunkhuman]

Benchmark

Run one freshly provisioned Knight from level 8 through level 20 at the tracked 1x progression rates. The normal controller must choose hunts, service, depot work, equipment acquisition, spell learning and use, and recovery from persisted player state. Do not author a level-by-level route or mutate progress to reach checkpoints.

Spell progression may require only normally acquired Thais spells: Find Person, Light, Light Healing, Cure Poison, and Great Light. Fixture-prelearned premium Haste and Whirlwind Throw do not satisfy benchmark spell progression until intercity trainer travel exists.

Required evidence

• Hunt cycles, pressure observations, frontier changes, selected regions, XP, damage, healing, retreats, and deaths.
• Loot sales, supply purchases, equipment spending, bank balance, depot value, and net cycle economics.
• At least one verified equipment improvement from an implemented deterministic source.
• Eligible spell candidates, training, casts, mana reserves, cooldown failures, and verified effects.
• Restart recovery during progression and normal save/reload at completion.
• Decision time, pathfinding calls, failures, suppressed telemetry, stuck events, and terminal outcomes.

Acceptance

• A documented reproducible command completes level 8-20 without fixture-only route checkpoints or direct gameplay mutation.
• The Knight completes repeated real Thais service, depot, and hunt cycles throughout the run.
• Hunt challenge escalates from the initial low-level region and backs off safely if observed pressure requires it.
• Equipment and spell progression occur through normal verified interactions when eligible opportunities exist.
• Supply reserves and economy remain viable; any external seed required by the benchmark is explicit and bounded.
• A deliberate server restart reconstructs a valid next objective from persistent state.
• No unresolved terminal, repeated deadlock, parser error, restart loop, or runaway scheduler behavior remains.
• The final report records elapsed time, XP/hour, deaths, potion and mana use, net gold, upgrades, spells, regions, and controller cost.

Benchmark result

Completed on 2026-08-20. The level-8 Knight reached level 20 and continued to a saved level 22 with 139,765 XP and 547 bank gold. The run took about 35 hours of wall time, including iterative rebuilds, validation pauses, and deliberate restarts, for an approximate wall-clock rate of 3,900 XP/hour; this is not an uninterrupted throughput measurement.

Observed progression included snakes, wolves, bears, wasps, trolls, troll champions, orcs, and rotworms. The bot completed repeated real Thais selling, potion restocking, banking, depot, equipment, spell-training, and hunt cycles. It equipped carried and purchased upgrades, learned Find Person, Light, Light Healing, and Cure Poison, and maintained the 10-potion target with a one-potion return threshold. One early dense-rotworm death recovered through normal lifecycle handling. Verified potion recovery occurred throughout; retained final segments recorded no verified Light Healing cast.

The longest retained controller segment recorded 128,598 decisions, 18,584 pathfinding calls, 9,569 pathfinding failures, 20 stuck events, about 1,004 seconds of decision time, and about 989 seconds of pathfinding time. Most excess cost came from the inaccessible-corpse loop that was later bounded by PR #114. Aggregate counters are segmented by rebuilds and restarts.

PR #114 bounded inaccessible corpse navigation and combat preemption. PR #115 independently bounded ordinary hunt-patrol plan failures. In the post-fix 20-minute wasp-cave comparison, pathfinding failures fell from 813 to 31, the prior 82-second stationary burst did not recur, and no death, restart, or terminal event occurred. The benchmark acceptance contract is complete.

Issue #111 separately tracks unsupported slotted-loot disposition; its immediate no-op service loop is mitigated and did not block benchmark completion.

A scope-exhausted result before level 20 is a benchmark failure with useful evidence, not permission to add intercity travel to this issue. Markets, other vocations, and multiple bots remain outside the milestone.

Related: #18, #33, #35, #38, #46, #51, #54, #57.

[adrunkhuman]

PR #114 merged as 2cd5c0e2b and bounded inaccessible corpse work. PR #115 merged as 1354a0805 and independently bounded ordinary patrol-plan failures.

The post-fix 20-minute wasp-cave comparison reduced pathfinding failures from 813 to 31. The prior 82-second stationary burst did not recur. One exhausted patrol destination ended after three calls, 1.384 seconds, and 230,574 expanded nodes. The run emitted no death, restart, or terminal event.

The final acceptance item is checked. The level 8-20 benchmark contract is complete; broader unsupported slotted-loot policy remains in #111 and does not reopen this benchmark.