docs: correct the Dispatchers.IO and publishing claims

[youndie]

Corrects two statements in the documentation. No code changes; one comment in publish.yml is updated.

Dispatchers.IO exists on Kotlin/Native

Research §1.8 said Dispatchers.IO is internal on Kotlin/Native. The member really is internal. The public Dispatchers.IO on Native is an extension property in kotlinx.coroutines, and without import kotlinx.coroutines.IO the compiler picks the member and reports "it is internal".

Evidence (macosArm64, Kotlin/Native 2.4.20, coroutines 1.11.0):

• klib dump-metadata of kotlinx-coroutines-core-macosArm64Main-1.11.0, the artefact §1.8 quoted, shows both declarations:

  internal final val IO: kotlinx/coroutines/CoroutineDispatcher
  public final val kotlinx/coroutines/Dispatchers.IO: kotlinx/coroutines/CoroutineDispatcher
  

• One file, compiled two ways. Without the import: cannot access 'val IO: CoroutineDispatcher': it is internal in 'kotlinx.coroutines.Dispatchers'. With the import it compiles, and withContext(Dispatchers.IO) runs on DefaultIoScheduler.
• kore B-42 found the same independently on linuxX64 and macosArm64, and measured that Native's IO dispatcher is elastic.

The research keeps the original text, adds a dated correction, and fixes the §1.4 row that had been struck out wrongly. CLAUDE.md, the README, feature-crud-mvp, mongkn-core and the M2 summary now say the corrected thing.

The design decision. MongoClient owns its thread pool, and the only reason ever recorded was this claim. Since then the pool's size has gained measured reasons (M-78, M-85), but nobody has weighed it against Dispatchers.IO.limitedParallelism(n). M-92 opens that question, with the arguments on both sides and the acceptance criteria. Code and KDoc (MongoClient.dispatcher, CollectionOps) are left for M-92, because the KDoc explains the design that M-92 will decide.

Publishing, as it works now

CLAUDE.md still described publishing.gradle.kts, which 3f6196c removed. It now describes the current setup:

• io.github.youndie.sborka.publish declares the wip repository and builds the POM; sborka.base sets the group.
• The convention reads REPOSILITE_USER / REPOSILITE_SECRET, which CI fills from the MONGKN_REPO_* secrets.
• The address is overridden with sborka.snapshotRepository. MONGKN_REPO_URL is read only by the verification step, not by the upload. That secret is not set today; if someone sets it, verification will check a different place than the one the artefact went to. The publish.yml comment that pointed at the removed file now says this too.
• The runner label is no longer copied into the prose. It lives in publish.yml, which says ubuntu-26.04 while the text still said 24.04.