Engine panic: "resolved-rules journal settlement ordinal overflow: UnknownNode(ActivatedMana(SettlementNodeOrdinal(6)))" (game_state.rs)

[mcbradd]

Assertion

thread 'ai-commander-driver' panicked at crates\engine\src\types\game_state.rs:16203:14:
resolved-rules journal settlement ordinal overflow: UnknownNode(ActivatedMana(SettlementNodeOrdinal(6)))

begin_activated_mana_journal_node (CR 605.4a region) unwraps resolved_rules_journal.begin_activated_mana(...). Here it came back UnknownNode(ActivatedMana(SettlementNodeOrdinal(6))) - the journal was asked to open an activated-mana node whose parent settlement ordinal it does not know - and the expect aborted the process (exit 101). Unchanged on main @ cbdfa042a: types/game_state.rs:19115 and :19133 carry the same .expect, so both the activated-mana and (adjacent) triggered-mana journal entry points share the failure mode.

The failing ordinal is 6, i.e. a fairly deep chain of nested mana activations, which fits the "nested activation records its active parent as the causal dependency" doc comment directly above the call.

Evidence

pod-lab headless 4-player Commander batch, 2000 games (two candidate decks x 1000), engine pin 1b2021e7b, card data 4587508ed14801cb, K=40 workers, 540 s/game, measurement mode. 48/2000 games (2.4%) exited 101 across 7 distinct assertion sites; 1 lands here. It did not appear at all in the equivalent 2000-game batch at the older pin d960fc8b3, so it is either new in that window or simply rare.

seed	run	tier	T/actions	wall	pod (P0 to P3)
92000363	plunder	Medium	T15/200	70.7 s	Blood Rites, Scrappy Survivors, Turtle Power!, Plunder the Graves (Meren of Clan Nel Toth)

Filed separately rather than rolled up because it is a distinct invariant with its own recorded seed and full saved stdout; low frequency, but a p0-class abort.

Repro shape

ai-commander <card-data-dir> --seed 92000363 --difficulty Medium --feed <4-deck-feed.json>

One game per process; panic = exit 101. The exact pod feed is reconstructable from the recorded seat ids - happy to attach it on request.

Replay caveat, stated plainly: seed replay is only faithful when the opponent pool is byte-identical to the recording run, and the pool has grown 105 -> 119 decks since. Two other seeds from this same run replay deterministically today (92000008 -> #6209, 92000014 -> #6822, exit 101 on every attempt); three longer-game seeds from the same run did not. This seed's replay is in flight and I will post the result here.

Sibling sites from the same run

#7134 (pending-replacement mana cost, 17x - most frequent), #6209 / #7103 (ChangeZone terminal member, 14x), #6822 / #7052 (draw bookkeeping, 8x), #6706 (exile-play ledger, 3x), #6867 (paused-child continuation, 3x), #4155 (announced entry at finalize, 2x), #7135 (buried reveal continuation, 2x at the older pin).

Evidence only - no fix attempted.

Claude Code

[mcbradd]

Full replay ledger (run today, sequentially, K=1, on pin 1b2021e7b)

seed	site it panicked at	recording pin	replay result today
92000008	effects/mod.rs:1349 ChangeZone	1b2021e7b	exit 101, panics (2/2 today; 4/4 including an earlier A/B in both execution modes)
92000014	effects/draw.rs:541 draw bookkeeping	1b2021e7b	exit 101, panics (2/2 today; 4/4 including that A/B)
92000280	mana_abilities.rs:1873	d960fc8b3	no panic (timeout at 540 s)
92000762	mana_abilities.rs:1873	d960fc8b3	no panic (stall)
93000243	mana_abilities.rs:1873	d960fc8b3	no panic (stall)
92000365	mana_abilities.rs:1932	1b2021e7b	no panic (win)
93000588	mana_abilities.rs:1932	1b2021e7b	no panic (win)
93000618	mana_abilities.rs:1932	1b2021e7b	no panic (stall)
92000363	game_state.rs:16203	1b2021e7b	no panic (timeout at 540 s)
92000232	casting_costs.rs:8810	1b2021e7b	no panic (win)
92000778	casting_costs.rs:8810	1b2021e7b	no panic (win)

What this does and does not mean. The two short panic games (T15-T21, ~40 s) replay perfectly; the eight longer ones (T27-T62) diverge. The pool has grown 105 -> 119 decks since the recording runs, so at least some pods resolve to regenerated decklists and no longer follow the recorded trajectory. A non-panicking replay is therefore evidence about trajectory reproduction, not about whether the defect is fixed - and the assertion text for every site here is still present verbatim on main @ cbdfa042a.

So: treat the seed tables as a recorded census of 86 aborts over 4000 games at two pins, with full saved stdout for each, rather than as live repro handles. If a maintainer wants a live handle for a specific site, the reliable route is to re-run the batch at a current build with a frozen pool - happy to do that and report seeds.

Claude Code

[mike-theDude]

Triage analysis

Clarified summary
Opening a nested mana journal node can reference an unknown parent settlement ordinal and panic.

Expected behavior
Activated and triggered mana journal nodes always attach to a live parent settlement or return a handled error.

Observed behavior
The journal rejects ActivatedMana(SettlementNodeOrdinal(6)) as unknown and a live expect aborts.

Reproduction status
The issue records a batch occurrence, seed/pod context, and replay limitations. Both adjacent assertion sites remain on current origin/main.

Investigation findings
The failure is low-frequency but localized to a shared activated/triggered mana provenance boundary.

Likely scope
Engine resolved-rules journal parentage for nested mana activations.

Acceptance criteria

• Nested mana nodes never reference an absent parent ordinal.
• Activated and triggered entry points share the corrected parent lookup.
• A deep activation-chain test covers ordinal rollover/settlement boundaries.

Related issues
The other batch panic sites are distinct invariants.

[matthewevans]

Fixed on main by #7282 (0c0ef3456d): activated and triggered mana journal creation now rejects stale parent nodes after journal retention/reset. The merged regression constructs the former deep stale-parent boundary and verifies both entry points start fresh rather than overflowing.