History cache-density knobs: decimation, incremental vacuum, blob compression

[AndrewCTF]

Summary

Makes the position-archive cache density tunable — store more replay per byte at
fixed disk. Three opt-in knobs, all OFF by default so the current on-disk
format and behaviour are unchanged. bash scripts/verify.sh green (1753 passed).

The knobs (Settings)

Setting	What it does	Lever
history_decimate_after_hours + history_decimate_stride	Thin each track older than N hours to one fix in every stride during the maintenance pass. Recent data stays full-fidelity; the first fix of an old segment is always kept so short tracks don't vanish.	Tiered resolution — old data goes coarse, so far more time-span fits the byte cap. Biggest density gain.
history_incremental_vacuum	INCREMENTAL auto_vacuum + a cheap PRAGMA incremental_vacuum each pass instead of a periodic full VACUUM. Falls back to full VACUUM on an existing store until its next VACUUM converts it.	Free-page reclaim — deleted pages return to the file continuously instead of accumulating between hourly passes.
history_compress_extra	zlib-compress the per-fix extra JSON blob (alt/squawk/… metadata). The blob is write-only today — no query reads it — so there's no read path to change.	Compression — a pure density win at write-CPU cost.

_encode_extra / _decode_extra handle the (de)compression and tolerate a mixed
column, so the flag can be flipped without a migration and a future reader has a
decode path.

Not included (deliberately)

• Coordinate quantization (lat/lon/track REAL → scaled INT) is the other
  big density lever, but it's a schema migration over the whole archive, not a
  runtime flag — a follow-up with its own migration, not a dead config knob here.
• Value-based eviction — for a uniform time-series the value metric is
  recency, which the existing time-window + byte-cap retention already implements.
  Decimation is the density knob; no separate setting needed.

Tradeoffs

Decimation loses old-data precision; compression and incremental vacuum add CPU.
That's the intended trade — density for compute/fidelity, not disk. All default
off, so nothing changes until an operator opts in.

Tests

• test_decimate_thins_old_tracks_keeps_recent — exact-count thinning, recent
  data untouched, and a short old track keeps its first fix (falsified: dropping
  the first-fix rule makes the short track vanish and the test fails).
• test_incremental_vacuum_mode_and_reclaim — a fresh DB comes up in INCREMENTAL
  mode and _reclaim runs without error.
• test_encode_decode_extra_roundtrip — both modes round-trip, and decode
  tolerates a mixed column and garbage.