perf: take general-purpose parsing and query costs off the interactive hot paths - #93

…e hot paths (#93)

Applies the idea behind uv's custom lockfile parser (astral-sh/uv#20648):
stop paying a general-purpose engine on a hot path with a known input shape.

- history: SearchByPrefix ran a GLOB + GROUP BY + MAX(timestamp) query on
  every keystroke, aggregating every historical occurrence of every match.
  Distinct successful commands now load once in the background at store
  open into a text-sorted in-memory slice; lookup is a binary search plus
  range scan, Store.Add updates the index in place, SQL stays the fallback
  until the load lands. In-memory SQLite DSNs cap the pool at one conn,
  since each pooled :memory: conn gets its own empty database.
- completion: built-in docker/hash-builtins TOML specs are compile-time
  constants; parse once per process under sync.OnceValue instead of on
  every handler-table rebuild.
- config: Load parsed config.toml twice in full, three times on recovery.
  Happy path now decodes only the agent key; recovery reuses its raw map.
- parser: map[string]any trace payloads were built per line for Emit to
  discard when tracing is off; now behind trace.Enabled guards. Emitted
  events are unchanged when tracing is on.

Benchmarks: linux/amd64, Intel Xeon 2.80GHz, go1.25.0, -benchmem. Committed
first in 8687faa so before/after run identical benchmark code.

  go test -run '^$' -bench . -benchmem \
    ./internal/history/ ./internal/completion/ \
    ./internal/config/ ./internal/parser/

  history.SearchByPrefix — per keystroke
  -----------------------------------------------------
  case                         before     after  speedup
  10k rows, dense prefix       845 µs    6.2 µs     136x
  10k rows, narrow prefix      453 µs    3.8 µs     120x
  10k rows, miss              32.9 µs     86 ns     383x
  100k rows, dense prefix     15.2 ms     57 µs     268x
  100k rows, narrow prefix     7.9 ms     40 µs     196x
  100k rows, miss             32.7 µs    107 ns     307x

  completion — startup, `completions reload`
  ----------------------------------------------------------------
  benchmark                    before     after  speedup     allocs
  builtinPluginSpecs           138 µs     63 ns   ~2200x   552 -> 1
  NewPluginCompleter           137 µs    1.6 µs      86x  561 -> 10

  config.Load — startup
  ----------------------------------------------------------------
  case                         before     after  speedup     allocs
  clean                       44.0 µs   39.2 µs    1.12x       -46%
  recovery                     154 µs    131 µs    1.18x       -21%

  parser.Parse — per dispatched line
  ----------------------------------------------------------------
  case                         before     after  speedup     allocs
  regular command              699 ns     59 ns    11.8x     7 -> 0
  ?? prefix                    854 ns     18 ns    46.7x     9 -> 0
  pipe | ??                   1295 ns     31 ns    42.0x    10 -> 0
  inline ??                   1074 ns     75 ns    14.4x    11 -> 0

The history before-numbers scale with table size (0.85 ms -> 15.2 ms from
10k to 100k rows, i.e. visible typing lag); the after-numbers scale only
with the number of distinct matching commands. config.Load is modest by
design — it runs once at startup, so the win is not parsing the same file
two-to-three times.

New tests cover the prefix index: live-vs-loaded recency merge, top-k
ordering, literal prefix matching, failed-command exclusion, SQL fallback
before load, and serving with the database closed. go test ./... passes;
touched packages pass with -race. golangci-lint v2.8.0: 0 issues. e2e is
excluded — it doesn't compile on the base commit either
(ScenarioMockTransport predates agent.Transport's AvailableModels).