fix(tiering): register spoke-namespace files; harden tier-drop rule

[xe-nvdk]

Summary

• the tiering scanner errored on every spoke-namespace file (one path level deeper than the standard layout), leaving hub-side spoke data invisible to tiering; partition paths are now parsed by their validated date tail, the same approach hub compaction adopted for the identical reason in Edge sync: hub-side compaction of spoke namespaces #619
• spoke files register under the query-visible split (database = spoke ID, measurement = spoke database), matching what FROM \"rocket-01\".telemetry resolves to, so registered metadata can never diverge from query-time tier assembly
• spoke files are gated OUT of hot-to-cold migration (follow-up medium(tiering): cold migration of spoke-namespace files needs receipt-aware handling #687): legacy spoke-synced daily files carry sync receipts that a migration delete would make confirmPresent forget, re-accepting duplicates (the Edge sync: HubIndex.Forget MUST be wired before delete_after_sync or hub-side spoke-namespace retention ships #611 hazard class); the gate is driven by the same parser that registers files, after review showed both a content-based heuristic (numeric spoke measurement names) and metadata-based reconstruction (legacy rows) misclassify
• multi-tier hardening: a verified-empty tier is dropped from a query only when another tier produced a positive pruning result, so spoke-shaped queries (whose generated paths sit one level shallow and existence-filter to empty) keep their full globs instead of losing a tier to a listing error

Process

Same pipeline as #681/#684/#686: adversarial plan review (which overturned both original load-bearing assumptions: the registration split and unconditional migration), independent implementation review (which caught the numeric-measurement gate bypass), licensed e2e before merge.

Verification

• unit + integration tests: spoke/plain/deep parse shapes incl. numeric spoke measurements, spoke register-but-not-migrate, combineTierPruneResults truth table, spoke-shaped pruning outcome
• full tiering/api/pruning/edgesync suites pass, including under TZ=America/Costa_Rica
• licensed e2e with real S3: spoke + plain layout scanned with 0 errors (previously every spoke path errored), migration moved only the plain daily file, spoke files stayed hot, FROM \"rocket-01\".telemetry on the tiered hub returns correct rows, and the migrated plain daily file reads back from cold

Refs #686. Spoke cold migration ships separately via #687.