Fix stranded-terminal recovery on a half-solved field

[wormeyman]

Two bugs in the same recovery path, found while investigating why big-list blueprint 227 stops planning under the Factorio 2.1 terminal offsets (#81).

1. The recovery ran on a mutated clone

GetBestSolution swaps context.CenterToTerminals for a per-strategy clone and mutates it, and does not put the originals back when a strategy fails. SelectBestSolution then called EliminateStrandedTerminals on whatever was left, so the recovery reasoned about a half-solved field rather than the real one.

Measured on blueprint 227:

before:  entry terminalLocations=47  ->  components [43, 3, 1], best covers 21/23 pumpjacks -> STRANDED
after:   entry terminalLocations=82  ->  components [76, 3, 3], best covers 23/23 pumpjacks -> ok

Two pumpjacks looked unreachable purely because the failed attempt had already pruned their terminals.

That 82 is not an internal number agreeing with itself. An independent flood fill over the raw pumpjack geometry counted 82 free terminal cells, which is what says the restored view is the true one.

2. EliminateStrandedTerminals kept the wrong component

It walked one component and, if that component did not cover every pumpjack, assumed it was a stranded pocket and eliminated it. That only holds when the component it happened to walk first is the minority one.

On blueprint 227 the first walk reached 43 of 47 terminals covering 21 of 23 pumpjacks, and the old code eliminated all 43 - throwing away the main network to keep a two-pumpjack pocket, stranding 19 pumpjacks in a single step.

It now enumerates every component and keeps the one serving the most pumpjacks, which makes the choice independent of where the walk started.

What each change costs, measured separately

Change	Snapshot churn
Component selection	None. The branch that got it wrong fires zero times across all 1147 big-list blueprints, and on its own it leaves every snapshot untouched.
Context restore	Two snapshots, both in AllowsBlueprintWithNonBlockingIsolatedArea - the test that exercises this path by name.

Both selected plans in those two snapshots get better:

CC-DT     effects 114 -> 116, beacons 62 -> 64, pipes 77 -> 78
CC-FLUTE  effects 116 -> 116, beacons 64 -> 64, pipes 81 -> 77

Selection ranks on effects, then fewest beacons, then fewest pipes. CC-DT wins on the primary criterion; CC-FLUTE ties and drops 4 pipes.

Score.HasExpectedScore is unchanged.

An ablation, so the two are not conflated

The restore fix alone makes blueprint 227 pass. So the component fix is defensive rather than load-bearing. It is included because it is wrong on its own terms and free to correct, not because 227 needs it.

Checks

• dotnet test: 4299 passed, 0 failed.
• dotnet build /p:UseLuaSettings=true: clean.
• src/lua regenerated and committed; tools/check-lua.sh passes both the Lua 5.2.4 syntax check and the sample.lua run.

Relationship to the other work

This is the second of two latent planner bugs that the 2.1 terminal offsets expose. The first is the Group.Location centroid collision in #87. Both are prerequisites for landing #81, and both are separated out so a subtle logic fix does not land inside a 2743-file snapshot diff.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P9FADuTnjE7SFEQWnpNhfc