core: expose running supplied cost in flow snapshots

[pzhin]

Summary

• expose exact runningSuppliedCost() for the current in-flight supplied service estimate
• expose derived completedSuppliedCost() as dispatched supplied cost minus the running gauge
• update the gauge atomically with dispatch and completion under the scheduler lock
• document that supplied costs are caller estimates, not actual execution time
• cover heterogeneous running costs, partial completion, reference equivalence, and snapshot/completion races

Verification

• ./mvnw --batch-mode --no-transfer-progress verify (re-run after rebase at 2519b7a)
  • 110 core tests and 5 integration tests passed
  • Checkstyle, JaCoCo, SpotBugs, JavaDoc, and artifact-surface gates passed
• ./mvnw --batch-mode --no-transfer-progress -Pbenchmarks clean verify
  • benchmark discovery and fixture smokes passed
• ./mvnw --batch-mode --no-transfer-progress -Pjcstress clean verify
  • 168/168 planned executions passed with no failed or error outcomes

JMH before/after

Environment: Apple M4 Pro, macOS Darwin 25.5.0, OpenJDK 21.0.11. Baseline is e49a621; the measured implementation is 18c18fd. Current rebased head 2519b7a has identical core source and tests; the rebase integrated current main documentation and CI changes, with the formal-spec conflict resolved in English. Each point used 3 forks, 5 x 1 s warmups, 8 x 1 s measurement iterations, and the GC profiler.

Workload	Before	After	Change
completeSteady	34.502 ± 0.615 ns/op	37.407 ± 0.673 ns/op	+8.42% latency
dispatchOne	249.353 ± 8.899 ns/op	243.457 ± 8.949 ns/op	-2.36%; intervals overlap
singleJobCycle, SKEWED_COSTS	1,920,717 ± 21,129 ops/s	1,796,462 ± 52,326 ops/s	-6.47% throughput
cycle allocation	2791.773 ± 0.003 B/op	2793.208 ± 0.003 B/op	+1.435 B/op

The cycle measures dispatch -> complete -> enqueue in steady state. These are machine-specific measurements, not a service-level claim.