[sparr]

Reduces an uncached launch of a large RP-1 install by about 55% to the main menu, 76% off ModuleManager's patching time. The patched output is identical. Launches with an not-invalidated cache are unaffected.

Commits

1. Speed up patch node matching (bb61d3a).
This commit can be broken up into smaller commits if requested.

• Currently, WildcardMatch builds a regex pattern string and a closure on every call before checking its cache, and runs a regex for patterns with no * or ?. This PR compares such names directly and caches regexes by pattern.
• Currently, CheckConstraints re-parses its constraint string for every node. This PR introduces a new parsed constraint class and caches after the initial parse of each constraint.
• This PR walks matching subnodes in one pass, instead of calling FindConfigNodeIn for each index. FindConfigNodeIn no longer builds a list.
• Generally, allocations removed by all of the above changes result in fewer GC waits on the patching thread.

2. Index root nodes by type for edit patches (904fc4b). Currently, every edit patch tests every root node, although a matcher only accepts one exact node type. This PR indexes nodes grouped by type, in list order, and edit patches visit only their own type's group. Edits replace values in place and cannot rename a root node, so they leave the index valid. Copy, delete, and insert patches mark the affected type stale, and it is rebuilt when next needed. IPatch.Apply is unchanged.

3. Match patches against each config's stored name value (ccec578). Currently, NodeMatcher calls GetValue("name") on every node of the patch's type, for every patch with a name pattern. This PR caches the name value when ProtoUrlConfig reads it. Renames already trigger another ProtoUrlConfig so that's automatically handled. This relies on nothing changing a root node's name value in place during patching, and I couldn't find anything that does.

Benchmarks

Time for a forced full repatch. "Patching" is ModuleManager's logged Ran in time. "Total" runs from launch to the main menu scene change.

Patches applied	Patching	Total
None	463 s	642 s
1: faster matching	204 s	387 s
2: type index	416 s	595 s
3: stored name value	416 s	592 s
1 + 2	131 s	307 s
1 + 2 + 3	111 s	285 s

• The install is KSP 1.12.5 with RP-1 and various other mods, 164,005 applied patches, on a Ryzen 5 5600X under Linux. It ran on a virtual display through VirtualGL, without an attached profiler or debugger.

Verification

• ModuleManager.ConfigCache, .TechTree, .Physics, and .ConfigSHA were identical to the unpatched build's in all six configurations.
• A temporary build of all three commits compared the stored and live name on every match during a full repatch and found no difference. The same check, given one deliberately wrong stored name, logged it and changed the output.
• The test suite passes under Mono: 476 tests, 14 of them new.

Behavior differences

• CheckConstraints stops at the first matching subnode. A later subnode with a null name would previously have thrown, and now is not reached. Null node names are not expected in practice.

AI Disclosure

Claude Opus 5.5 produced 99% of the code, 100% of the commit messages, and 50% of this PR description. I have lightly reviewed the code and am proceeding with a more thorough review. I will un-draft this PR when I am confident in its implementation. I won't be offended at all if this is considered an issue with POC code instead of a PR and someone reimplements it by hand instead.

[sarbian]

Nice. Those numbers looks great. I will let you test it.

[sarbian]

One correctness issue in the third commit (ccec578): the assumption that nothing changes a root node's name value in place doesn't hold.

The Special * command can write to another root node directly. RecurseVariableSearch returns the live ConfigNode.Value from the database config, and ModifyNode assigns to it (val.value = modVal.value) without creating a new ProtoUrlConfig, so NameValue keeps the old name.

@PART[source] { *@PART[target]/name = renamed }
@PART[renamed] { x = 1 }
@PART[target] { y = 2 }

On master, @PART[renamed] applies to that node and @PART[target] doesn't. With this PR it's the reverse, with no error logged, until another edit happens to rewrap the node. The temporary stored-vs-live check wouldn't catch this unless the install has such a patch, and no test covers *@; TestIsMatch__ProtoUrlConfig__UsesNameValue actually pins the stale behaviour.

Possible fixes: have the Special branch refresh the affected ProtoUrlConfig, or fall back to the live name when a patch run contains a *@ write that targets a root node.

Found by reading the code; I haven't run this scenario.