[codex] perf: speed up warm snapshot reopen

[justrach]

Summary

• restore snapshot outlines/state directly instead of reparsing every file on warm reopen
• reuse persisted trigram sidecars for CLI and MCP loads, and lazily rebuild/persist the word index
• skip redundant word.index rewrites and align the clean main-based branch with AnyTrigramIndex

Why

Warm snapshot startup was still spending seconds rebuilding indexes after loading codedb.snapshot. This keeps cold indexing effectively flat while making warm CLI and MCP startup dramatically faster.

Benchmarks

All numbers below are ReleaseFast and measured against v0.2.54 on the same machine. The broader benchmark appendix was added post-merge while preparing the 0.2.55 release.

Data Shape: Large Repo CLI (openclaw)

Benchmark	0.2.55	0.2.54	Delta
cold tree	5.32s	5.29s	+0.6%
snapshot	6.53s	6.25s	+4.6%
warm tree	0.26s	6.16s	23.7x faster
warm search workspace	0.24s	6.14s	25.6x faster
warm word session	0.61s	5.99s	9.9x faster

Data Shape: MCP First Secondary-Project Call (openclaw)

Tool	0.2.55	0.2.54	Delta
codedb_tree	0.076s	5.289s	69.6x faster
codedb_search	0.067s	5.278s	78.8x faster
codedb_word	0.285s	5.312s	18.6x faster

Data Shape: Peak RSS (openclaw)

Benchmark	0.2.55	0.2.54
cold tree	3478.8MB	3478.1MB
warm tree	192.6MB	3314.0MB
warm search	193.3MB	3312.9MB
warm word	677.1MB	3313.3MB

Data Shape: Small-Corpus Sanity (codedb/src)

Benchmark	0.2.55	0.2.54
cold tree	0.045s	0.040s
warm tree	0.010s	0.030s
warm search	0.010s	0.030s
warm word	0.010s	0.030s

Workload-Shaping Note

The release also includes WAL profiling and hashed cloud sync from #198 and #199, so production query/open patterns can be aggregated without sending raw query strings or file paths off-machine.

Validation

• SDKROOT=$(xcrun --show-sdk-path) zig build test
• SDKROOT=$(xcrun --show-sdk-path) zig build -Doptimize=ReleaseFast
• SDKROOT=$(xcrun --show-sdk-path) zig build run -- --help

Closes #220.