feat(spartan): client-observed inclusion latency for bench-10tps

[spypsy]

Summary

Replaces the Prometheus aztec_mempool_tx_mined_delay_milliseconds histogram (which mixes every mempool observer — RPC + validators + full-nodes — and both high- and low-value txs) with a client-observed measurement scoped to the high-value lane.

What changes

• n_tps.test.ts: block-watcher RunningPromise polls getBlockNumber() every 1s and stamps wall-clock minedAtMs the first time each sent tx's block becomes visible to the test client. Records dumped (high-value group only) into /tmp/n_tps_timing_data.json alongside the existing startedAt/endedAt.
• tx_metrics.ts: rename sentAt/minedAt/attestedAt → *Ms (ms wall-clock); add observeBlockForMinedTxs (idempotent, first observer wins) and getInclusionRecords. recordMinedTx kept as fallback (slot-timestamp × 1000) for any tx the watcher misses.
• bench_scrape.ts: load records via new --inclusion-records flag; compute summary.inclusionLatencyP{50,95,99}Ms and the txMinedDelayP{50,95,99} time series (60s bins by sentAtMs) directly from the per-tx data. The other Prom-based histograms (build duration, public-processor) are untouched.
• bench_output.schema.json: widens timeSeries.source enum to include "client_observed", declares the three txMinedDelay* slugs.
• bootstrap.sh: HIGH_VALUE_TPS=1, LOW_VALUE_TPS=9 (was 10/0) so the high-value scoping is non-trivial; passes --inclusion-records "$metadata" to the scraper.

Field names on the run JSON are unchanged, so the dashboard panel in AztecProtocol/explorations keeps working — descriptions updated to reflect the new source (PR spy/inclusion-latency-client-source).

Why

• Prom histogram measures mempool-add → mempool-saw-block per node. Aggregating across RPC + validators dilutes the headline with the validator's gossip-arrival-based view.
• It also doesn't distinguish high-value from low-value lanes. In the new 1/9 mix, low-value txs run at network min priority and are explicitly allowed to fail fee checks, which would dominate any aggregated quantile.
• Client-observed wall-clock is what a paying client actually experiences end-to-end.

Caveats

• minedAtMs has ~1s jitter from the watcher's poll interval. p50 is typically multiple slots, so this is small relative to the signal.
• bench_output.schema.json diff is noisy: the precommit prettier hook expanded inline { "$ref": "..." } entries throughout. Semantic change is ~7 lines at the top of the file.

Test plan

• Run bench-10tps end-to-end against a real namespace.
• Verify /tmp/n_tps_timing_data.json contains inclusionRecords with non-trivial minedAtMs - sentAtMs deltas.
• Verify run JSON has timeSeries.txMinedDelayP* with source: "client_observed" and summary.inclusionLatencyP*Ms populated.
• Verify the dashboard panel renders the line chart on the new run.

Supersedes #23465.