Working state for the tgf mill mine / tgf mill mine-endgame pipeline that produces src/ui/flutter_app/assets/patches/std.mill_patch (see crates/perfect-db/src/patch.rs and crates/tgf-cli/src/mill_mine/, mill_endgame/, mill_pack/, mill_arena/ for the code).

Everything under this directory is gitignored (see the root .gitignore's /mining/*/ rule) except this file: entries.jsonl and checkpoint.json are large, fully regenerable-from-the-database working data, not source. They are kept on disk (not in git) purely so mining can resume from where it left off instead of restarting from scratch.

All runs below used the external full Malom database (not the small subset bundled under src/ui/flutter_app/assets/databases/) at D:/user/Documents/strong and D:/Repo/NMM_LLM/human_database/human_db.sqlite for human-seeded runs. Adjust --db/--human-db if those paths differ on the machine that resumes this.

Each holds entries.jsonl (mined MineEntry records, one per line -- the actual training data), checkpoint.json (resumable frontier/visited state), and log*.txt (eprintln! output from the run(s) that produced them, for reference).

• opening/ -- mill mine from the empty board, --max-depth-plies 10 --budget-engine-calls 20000. Fast, low remaining value; only worth resuming if opening coverage regresses.
• placing/ -- mill mine --human-db ... --placing-only --root-mass 0 --budget-engine-calls 15000. Superseded in practice by placing_expanded/ below (same seeding, much larger budget); kept as-is since it was one of the inputs to the currently-bundled patch.
• placing_expanded/ -- same as placing/ but --seed-phase placing --budget-engine-calls 400000 --budget-seconds 7200. The highest-value resume target: placing-phase coverage is still thin relative to its ~54-61% blunder density, and checkpoint.json's remaining_frontier was in the hundreds of thousands when this last stopped on the time budget, not exhaustion.
• moving/ -- mill mine --human-db ... --seed-phase moving --max-depth-plies 8 --root-mass 0, run across several --resume passes with growing budgets. checkpoint.json is ~1.6GB; expect a --resume run to need a similar amount of free memory to load it.
• endgame/ -- mill mine-endgame --min-side-pieces 3 --min-total-pieces 6, exhaustive (not sampled). checkpoint.json here is just the list of fully-completed (white_on_board, black_on_board) sectors (see mill_endgame's own checkpoint format), currently total-pieces 6 through 9. Resuming with a higher --max-total-pieces picks up new sectors only; already-completed ones are skipped. Sector density collapses fast past 4v4-ish material (measured ~0-1% at total=8/9 vs ~38-60% at total<=7) -- expanding this further is low priority; see placing_expanded/ instead.
• closed_loop/ -- mill mine --seed-fen-file seeds.txt --root-mass 0 --max-depth-plies 24, seeded from mill arena --out's first_uncovered_blunder_fen values (a prior run's actual, still-losing games). seeds.txt is the seed list used; regenerate a fresh one from a new mill arena --out arena.jsonl run (see that command's module docs) before resuming this if you want it to target today's remaining losses rather than the ones already fixed.

Same flags as the original invocation, plus --resume and the same --checkpoint/--out paths, e.g.:

tgf mill mine \
  --db "D:/user/Documents/strong" \
  --human-db "D:/Repo/NMM_LLM/human_database/human_db.sqlite" \
  --seed-phase placing --placing-only --root-mass 0 \
  --out mining/placing_expanded/entries.jsonl \
  --checkpoint mining/placing_expanded/checkpoint.json \
  --resume \
  --workers 20 --budget-engine-calls 400000 --budget-seconds 7200

mill mine-endgame resumes automatically (it always skips sectors already listed in --checkpoint); just raise --max-total-pieces and rerun.

tgf mill patch-pack \
  --in "mining/opening/entries.jsonl,mining/placing/entries.jsonl,\
mining/placing_expanded/entries.jsonl,mining/moving/entries.jsonl,\
mining/endgame/entries.jsonl,mining/closed_loop/entries.jsonl" \
  --db "D:/user/Documents/strong" \
  --out src/ui/flutter_app/assets/patches/std.mill_patch \
  --budget-bytes 60000000 \
  --recompute-from-fen \
  --audit-sample 10000

Then validate with tgf mill arena --db "D:/user/Documents/strong" \ --patch src/ui/flutter_app/assets/patches/std.mill_patch --games 1 \ --max-plies 200 --out arena.jsonl before committing the regenerated asset -- compare losses/draws against the previous run, and check arena.jsonl's first_uncovered_blunder_fen for new closed_loop/ seeds.