add fs ephemeral and txoutset build logging

[idky137]

Finalised-state observability

Additive observability only — no control flow, status semantics, or routing behaviour changes.

Why

Two blind spots hit while syncing a mainnet node:

1. The accumulator rebuild is silent. rebuild_tx_out_set_accumulator logs two lines up front and nothing again until it commits. In between it runs a full spent count pass and one full-chain block scan per shard — ~40 minutes on a mainnet DB — with no output. "Working" and "hung" look identical.

2. chain_state: Ready doesn't mean the persistent DB is serving. During a long sync the router installs an ephemeral passthrough; it tracks the validator and reports Ready exactly like a synced on-disk index. Nothing logged the install or release, and nothing exposed routing state. Containerised tests waiting on Ready have been silently asserting against the passthrough.

What

Rebuild progress — spent-entry count pass, per-shard start/spent-set-size/completion, and intra-shard height progress. Throttled to one line per 10s (reusing the existing write_core idiom, now a shared PROGRESS_LOG_INTERVAL) so output stays bounded regardless of shard count — a memory-constrained container can reach 256 shards. Also covers the startup spent integrity check and the incremental update path.

A shard bisect logs at WARN, naming accumulator_rebuild_memory_size — it means the budget under-provisioned and each split adds another full-chain pass.

Routing mode — new FinalisedStateMode (EphemeralConfigured / EphemeralRouted / Persistent), reported on every transition plus a one-shot "finalised state online" line:

INFO router: finalised state switched to the ephemeral passthrough: reads are served from the
backing validator, NOT the persistent database ... mode=ephemeral(syncing)
INFO router: finalised state switched back to the persistent database mode=persistent
INFO router: finalised state online: reads are now served by the persistent database

The one-shot latch fires from both the ephemeral-release edge and the status poll — a restart against an already-current DB never installs a passthrough, so the release edge alone would never fire.

ephemeral_finalised_state = true now warns at startup; that branch previously returned completely silently.

Status line gains fs_mode, so CI can gate on fs_mode: persistent instead of Ready.

Metrics (prometheus feature, off by default): zaino.db.finalised_ephemeral, zaino.db.accumulator_built_height, zaino.db.accumulator_rebuild_active.

Notes for review

• fs_height was dropped from the status line: db_height() is async, log_status is a sync pub fn, and making it async is a breaking change. The height is available via the gauges.
• One private-fn signature change (shard_ordinal/shard_total for logging), one pub accessor chain mirroring the existing status() chain, one AtomicBool latch on Router.
• Drive-by fix: zaino.chain.tip_height was emitted twice per sync iteration, once via a hard-coded string literal and once via CHAIN_TIP_HEIGHT. Removed the literal. Happy to split this out.
• makers lint fails on this branch before this PR — --all-features enables transparent_address_history_experimental, which doesn't compile (write_core.rs:404, :1541). Verified against a clean tree. Not addressed here.

Verification

cargo fmt --check, cargo clippy --workspace --all-targets and --features prometheus all clean (no new warnings); makers lint-boundary-conversions clean; 586 tests pass. Two new tests cover the configured-ephemeral mode and the install→release transition.

14 files, +584/−19.

Out of scope and worth separate issues: checkpointing the rebuild so restarts resume, relaxing NO_READAHEAD for the rebuild's sequential scans, and whether Router::status() should report Syncing while ephemeral-backed.