feat: streaming transcript/checkpoint parsing to reduce memory amplification

[Krishnachaitanyakc]

Summary

• Convert 4 JSONL transcript parsers (Claude, Codex, Windsurf, Droid) from read_to_string to streaming BufReader + .lines(), eliminating full-file memory materialization
• Convert 2 JSON parsers (Gemini, Continue) from read_to_string + from_str to from_reader, avoiding intermediate String allocation
• Stream checkpoint JSONL reads via BufReader instead of read_to_string
• Eliminate 2 redundant read_all_checkpoints() calls in get_all_tracked_files() by reading once upfront and threading as &[Checkpoint]
• Perform hash migration in-place (&mut iteration) instead of allocating a second Vec
• Add configurable size caps (max_checkpoint_jsonl_bytes=64MB, max_transcript_bytes=32MB) via env vars or file config; checkpoint cap is advisory-only, transcript cap returns Err to preserve existing data
• Handle --reset with eager reset on read failure, propagating write errors independently
• Harden generated CI workflows with action version pinning and integrity checks

Motivation

git-ai's transcript/checkpoint parsers cause ~5-8x memory amplification: a 187 MB transcript produces ~1.2 GB peak RSS, and a 307 MB checkpoint file produces ~1.78 GB peak RSS with ~33s runtime. This causes OOM kills and poor UX on long AI coding sessions.

Targets

• Transcript peak RSS down ≥40%
• Checkpoint peak RSS down ≥30%, wall-clock down ≥25%

Test plan

• cargo fmt -- --check — clean
• cargo clippy --all-targets -- -D warnings — clean
• cargo test --lib — 1205 passed, 0 failed
• Run scripts/repro_runaway_memory.py for before/after RSS measurements
• Manual testing with large transcript files (>100MB)
• Verify --reset recovers from corrupt checkpoints.jsonl