fullhistory: simplify LastCommittedLedger to the design signature LastCommittedLedger(cat)

[chowbao]

Follow-up from PR #820 review (thread on backfill/process.go, comment 3515787395).

The nil-logger positional mode of lifecycle.LastCommittedLedger(cat, logger) is production-dead (the only production caller, startup.go, passes a real logger; the tick's LastCommittedLedger(cat, nil) went away with the #25 clamp deletion). It should be removed, along with the logger parameter, landing the design-pseudocode signature LastCommittedLedger(cat).

Blocked on two coupled changes, which is why it wasn't folded into the review round:

1. The TestLastCommittedLedger unit tests use the nil-logger mode as a seam to exercise the positional derivation WITHOUT opening a real hot DB. They must be converted to real-DB tests (as TestDeriveWatermark_RealHotDB_RefinementIsNotStale already is) or the positional-only cases dropped.
2. The refinement's read-only open needs a logger; drop the param by defaulting nil->nop in rocksdb.Config (the same idiom orNop applies to MetricSink), or retain a logger on Catalog.

Pure cleanup — no behavior change.

[tamirms]

With the test-seam half folded into d2b1c127, the only blocker left here is the logger plumbing — small enough that it may as well finish in #820 rather than wait. Of the two options, I'd take the logger-on-Catalog shape: the nil→nop default in rocksdb.Config would silently drop diagnostics on the crash-recovery read path (the derivation's read-only open is the WAL-replay read the resume point depends on), and it reintroduces the nil-tolerant logger mode this review just removed. Catalog is built once in the daemon with a real logger, so derivations get logging without the parameter, and LastCommittedLedger(cat) lands as designed.

One more observation that sharpens the fold-in: since the tick stopped calling it (the clamp deletion), nothing inside the lifecycle package uses any symbol in progress.go — its only production caller is startup.go. So the function can move next to its caller in the fullhistory root and drop the export entirely, landing as lowercase lastCommittedLedger(cat) — the design pseudocode's exact name and shape — which also settles the "where does progress live" half of the #824 placement question for this file.

[tamirms]

Post-merge freshness note: blocker 1 in the body (converting the nil-logger positional unit tests to real-DB) landed in #820 (d2b1c127 — the tests now seed real empty hot DBs). The remaining work is only blocker 2 — the logger plumbing (logger on Catalog preferred over a nil→nop default in rocksdb.Config, which would silently drop diagnostics on the crash-recovery read) — plus moving progress.go next to its only caller in the fullhistory root and dropping the export, per the scope recorded from the #820 thread.

[chowbao]

Scope update from PR #844 (commit 8fa6c1a): progress.go moved to the fullhistory root next to its only production caller (startup.go) and LastCommittedLedger is now the unexported lastCommittedLedger. What remains here is just the logger-parameter change — dropping the param by defaulting nil→nop in rocksdb.Config (or retaining a logger on Catalog). The nil-logger test seam is already gone: the unit tests pass a real silent logger and the positional cases run against real (empty) hot DBs.

[chowbao]

Done in PR #844 @ a051a06: the metastore→catalog fold gives the catalog a logger, refineWithHotDB uses cat.Logger(), the logger parameter is dropped, and lastCommittedLedger(cat) lands as the design wrote it. The PR now carries Closes #835 and will close this on merge.