Fix PRISM latest-tip refresh supersession

[Anatolie]

Problem

During rapid tip changes, a canceled PRISM bundle build kept logical refresh ownership until its executor physically exited. A newer tip was selected correctly but could not start, so repeated B → C → D changes left miners on stale work while B unwound.

What changed

• Separate logical cancellation from physical executor cleanup. Superseded waiters complete immediately while the old build retires in the background.
• Admit the newest replacement without waiting for one obsolete executor, while keeping a hard bound of two physical builds and one latest pending request.
• Replace intermediate pending generations so B → C → D publishes D rather than C.
• Carry generation-bound cancellation through payout preparation, expensive serialization and copy steps, scheduler admission, and result handoff.
• Fence promise completion and cache publication so a retired build's late result cannot publish, populate cache state, or satisfy a newer request.
• Preserve refresh retry ownership, pending tokens, block-candidate retries, PR Isolate PRISM share hot-path locking #83 lock separation, and existing timeout values.

Expected outcome

• The latest tip can begin and publish while obsolete work is still exiting.
• Current-tip job coverage should recover quickly during short reorgs or rapid block bursts, reducing stale and unknown-job rejects.
• Repeated tip churn remains bounded and cannot create an executor or promise storm.

Validation

• 192 focused refresh, cache, and publication tests passed.
• 179 PR Isolate PRISM share hot-path locking #83 hot-path, share-ledger, initial-job, and reconnect tests passed.
• 275 Vardiff and block-candidate retry tests passed.
• Full Python suite: 1,231 tests passed.
• Python compilation, Ruff, shell syntax, ShellCheck, and diff checks passed.

Monitor after merge

• Current-tip coverage and stale/unknown-job reject rates.
• Refresh first-delivery, last-delivery, replacement-start, and cancellation duration.
• Active/pending builds, supersession and obsolete-result counts, and builder terminations or restarts.
• Watchdog exits, candidate backlog/retries, and coordinator lock-wait latency to confirm the existing hot-path behavior remains stable.

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}