[GGGP M2] Archimedes strategy_spec Target adapter (first DSL)

[mnemonik-dev]

Summary

Milestone M2: the first concrete Target — the Archimedes strategy_spec DSL — used to prove the
whole pipeline against a real validity oracle. Implements the DSL grammar (BNF), the derivation→spec
builder, semantic repair, a Rust validity check mirroring validate_strategy_spec, and the
enum-conformance guard (D6) so the grammar can't drift from strategy_dsl.py. Fitness is
synthetic/Rust in this phase (the real backtrader + rigor-gate fitness arrives with the later
Python-bridge phase, D7).

Spec: https://github.com/mnemonik-dev/genetic_algorithms/blob/main/docs/technical-reference/gggp-implementation-spec.md (§5)
DSL source of truth: Archimedes backend/archimedes/services/strategy_dsl.py. BNF: gggp-grammar-tech-review.md §3.

Scope

• src/targets/archimedes.rs implementing Target:
  • grammar() — the DSL CFG: indicators {sma,ema,rsi,realized_vol,momentum}, cmp {gt,lt,gte,lte},
    logic {and,or,not}, rebalance {daily,weekly,monthly}, sizing
    {full_invested_when_in_market, equal_weight, inverse_vol, volatility_target}, price operands,
    period Int[1,10000], annual_pct Float(0,1], tickers via FromSet.
  • build() — derivation → exact strategy_spec JSON; inject source_arxiv_ids from
    GrammarConfig.metadata (provenance carried, not searched); set look_ahead_safe = true.
  • repair() — clamp period→[1,10000], ensure ≥1 asset, annual_pct>0, parameter_variants len∈[2,8], dedupe.
  • validate() — Rust port of the structural checks in validate_strategy_spec.
• examples/evolve_faber.rs — evolve with the Archimedes target + a synthetic fitness.
• Tests: validity-invariant, mapping-roundtrip (Faber), repair, enum-conformance.

Acceptance criteria

• cargo test -p evoforge-grammar → 0 failed.
• Validity invariant: 5_000 random + 5_000 evolved genomes → every built artifact passes
  ArchimedesTarget::validate (100%, 0 failures).
• Mapping roundtrip: hand-written codons reproduce FABER_2007_SPEC exactly.
• Repair: injected out-of-bounds period / empty asset_universe / annual_pct<=0 are repaired
  to specs that then validate.
• Enum conformance: a test asserts the grammar's terminal sets equal the DSL's frozensets
  (INDICATOR_NAMES, COMPARISON_OPS, LOGIC_OPS, REBALANCE_FREQUENCIES, POSITION_SIZING_TYPES).
  The expected values are pinned in the test (with a comment pointing at strategy_dsl.py); the test
  fails if the grammar diverges. (Since Python isn't imported here, the DSL values are mirrored as
  constants — divergence from the real file is caught at review against the cited source.)
• Convergence on the synthetic fitness; cargo fmt --check + clippy -D warnings clean.

Verify

cargo test -p evoforge-grammar
cargo run -p evoforge-grammar --example evolve_faber

Anti-goals (do NOT)

• Do not call Python or import the real backtest/rigor gate (D7 → later phase); fitness is synthetic.
• Do not widen the DSL's indicator/operator set (that touches the live Archimedes interpreter — separate, review-gated change).
• Do not add NSGA-II (D5), one-point crossover (D2 → M3), resume (D4), or PyO3 (D7).
• Do not modify evoforge core or strategy_dsl.py.

Notes

Depends on M1 (#2). After this, the GE pipeline is end-to-end provable in pure Rust; M3 (operators)
and M4 (release) follow, then the separate Python-bridge phase wires the real Archimedes fitness.

[mnemonik-dev]

Implemented on branch claude/brave-davinci-syla39 (commit 76e7e02) as evoforge-grammar/src/targets/archimedes.rs.

Acceptance — all met:

• cargo test -p evoforge-grammar → 18 passed, 0 failed (5 new in tests/archimedes.rs)
• Validity invariant ✓ — validity_invariant_random_and_evolved: 5,000 random genomes (every one that decodes builds and passes validate); evolved population checked across 25 generations — ask() never yields an invalid artifact (100%).
• Mapping round-trip ✓ — faber_mapping_roundtrip: hand codons [0,0,2,0,0,0,0,1,0,199,0,1,0,0,1,0,199,0] → exact FABER_2007_SPEC (entry {"gt":["close","sma_200"]}, exit {"lt":[...]}, monthly, full-invested, source_arxiv_ids:["0706.1497"], look_ahead_safe:true).
• Repair ✓ — repair_fixes_semantic_bounds: empty universe→["SPY"], sma_99999→sma_10000, annual_pct:-1.0→0.15; invalid-before / valid-after.
• Enum conformance (D6) ✓ — enum_conformance_with_dsl: grammar terminals == the DSL frozensets (INDICATOR_NAMES/COMPARISON_OPS/LOGIC_OPS/PRICE_OPERANDS/REBALANCE_FREQUENCIES/POSITION_SIZING_TYPES) + ticker set.
• Convergence ✓ — convergence_on_synthetic_fitness; cargo run --example evolve_faber reaches fitness 3.0 with a valid spec.
• cargo fmt --check ✓ · cargo clippy --all-targets -- -D warnings ✓

build/repair/validate are a Rust port of strategy_dsl.py's structural checks; provenance (source_arxiv_ids) is carried, not searched. Fitness is synthetic — the real backtrader + rigor-gate fitness (with DSR trial-count deflation) is the later Python-bridge phase (D7), not this milestone. Left open for review/PR.

---

Generated by Claude Code