Read the local Syzygy directory from TABLEBASE_PATH so a dev machine and a VPS can point at their own copies. Add mix tablebase.fetch (downloads the 3/4/5 WDL+DTZ set, resumable) and mix tablebase.check (verifies it loads), plus a boot log reporting whether local tablebases are ready and the largest piece count.
Covers the codeable part of #31; installing on the VPS and setting TABLEBASE_PATH there remains a manual step.
Add Backend.Fathom: probes positions with five or fewer pieces via the Fathom NIF, maps WDL/DTZ to the trainer structs (flipping per-move perspective), generates SAN, sorts moves best-first, and falls back to Lichess on any failure. The local directory is configured with TABLEBASE_PATH; local answers are no longer cached.
Add Tablebase.normalize_wdl/1 and flip/1, plus tests for the WDL mapping, SAN, checkmate, fallback, and a cache-free end-to-end local probe.
Vendor the MIT-licensed Fathom C source and compile it into a Rustler NIF (native/fathom_nif). The NIF exposes init/1 (point Fathom at a tablebase directory) and probe/1 (raw WDL/DTZ/move results for a FEN); probes are serialised because Fathom's root probe is not thread safe, and run as dirty NIFs.
Add KRvK/KQvK 3-piece WDL+DTZ fixtures and an integration test proving a known KRvK probe returns Win, DTZ 23 and 15 moves.
Add FEN.position_key/1 (board, side, castling and en-passant; clocks dropped) and have Tablebase.new/1 key cache reads and writes on it, so transpositions and clock-only differences share an entry. The Lichess category is a position property, so dropping the clocks is safe.
Update the cache docs and tests, and give the play fixture's after_first/after_losing distinct positions since they previously differed only by the halfmove clock.
Add a Backend behaviour, a Lichess implementation, and a Router that picks a backend by piece count: five or fewer to the local slot, six or seven to the remote slot, and eight or more short-circuited as not covered with no network call. Both slots point at Lichess for now; the local Syzygy backend replaces the local slot when it lands.
Map :not_covered to the existing 'not covered by the tablebase' message.
Drop dtm/dtc from Tablebase and Tablebase.Move; the local Syzygy backend cannot produce them and nothing reads them. The Lichess parser now maps only category, checkmate and moves (with per-move category, dtz, san and uci), and the struct-shape tests follow.
* Slim the cached tablebase value to the fields the trainer reads (#20)
* Sample for LFU eviction instead of scanning the cache (#21)
* Make cache size configurable and back up daily off-peak (#22)
* Document the cache design in the glossary and an ADR (#23)
* Address cache-scaling code review findings
#7: prefetch the tablebase at mount and grade a played move locally
against its moves list (optimal / winning / losing), blocking the
attempt when the tablebase is rate-limited or unavailable. The client
now reports prev_fen and promotion, and the trainer board locks after a
graded move until reset.
Fix the 429 clause ordering in the tablebase parser so the cooldown
actually fires, and use the HTTPS endpoint.
Add dialyxir to the dev/test deps.
#5: normalise the full lichess tablebase category vocabulary (cursed-win
and blessed-loss classify as draws), return {:ok, tablebase} on a cache
hit, and unwrap the tablebase in game_from_fen.
#6: replace the LRU cache with an LFU cache storing
{key, value, frequency, last_access}, evicting the least-frequently-used
entry (LRU tie-break) and halving every frequency on a weekly sweep.
Persistence moves to a versioned v2 dets file, dropping the legacy
backup; the cache restores synchronously at startup.
Also updates stale tags enum and Play LiveView tests, and drops the
tracked legacy cache backup.