feat(benchmark): dimension-gated arm activation — NLSHADE_LBC enters Rewarding_Restart at d≥5

[haraldschilly]

codify-slot: NLSHADE_LBC/add_heuristic (shipped in its gated form; the 2026-08-11 rejection was "rejected as an unconditional add")

What

Daily improvement session (GOAL.md §3/§4). Today's codify scan surfaced 0 actionable candidates (13 total: 5 already codified, 8 rejection-suppressed; last night's two accepts reinforce already-codified slots), so this session executes the §4 attack that PR #298 queued: dimension-gated arm activation (GOAL §5.1(d)).

• StrategySpec.create_strategy now honours two reserved heuristic-kwargs keys, gate_min_dim / gate_max_dim: the arm is only instantiated when gate_min_dim <= problem.dim <= gate_max_dim (either bound optional). The keys are stripped before the heuristic constructor sees them; curated MutationRules name (class, param) explicitly, so the reserved keys cannot be accidentally mutated by the nightly loop.
• First user: (NLSHADE_LBC, {"NP_init": "auto", "k_rank": 3.0, "gate_min_dim": 5}) in Rewarding_Restart — the exact structural-catalog kwargs NLSHADE_LBC add to Rewarding_Restart: REJECTED unconditionally — first CI-significant per-dim split (d5 gain, d2 loss); log entry + revert #298 measured. At d<5 the spec is arm-for-arm identical to master.
• Unit tests pin: gate below threshold, admit at threshold, key stripping, gate_max_dim, no mutation of the spec's kwargs dict.

Evidence

Why gated — #298's 12-seed standard A/B measured the unconditional add at d2 −0.0241 [−0.0401, −0.0080] vs d5 +0.0080 [+0.0007, +0.0154]: both CIs exclude zero, in opposite directions. The scalar objective could not express the gain; the gate ships it without the d2 loss.

Ledger nights — add_heuristic NLSHADE_LBC: 3 confirmed accepts across 2 nights (2026-07-09, 2026-08-10), pooled CI95 [+0.0075, +0.0192] — all on the RoundRobin_Random control, which is why no codify slot could reach the candidate spec and a direct A/B is the vehicle.

Local paired A/B (this PR) — 12-seed canonical roster, --standard, --sync-eval both sides, before = master 92b603e:

strategy	dim	before	after	Δmean	CI95	verdict
Rewarding_Restart	d2	0.4330	0.4297	−0.0033	[−0.0131, +0.0065]	flat (gate off — spec identical to master)
Rewarding_Restart	d5	0.3026	0.3088	+0.0062	[−0.0001, +0.0125]	positive, boundary
RoundRobin_Random (control)	d2	0.3618	0.3613	−0.0005	[−0.0019, +0.0009]	flat
RoundRobin_Random (control)	d5	0.2834	0.2846	+0.0011	[−0.0052, +0.0074]	flat

At d5 the gated spec is byte-identical to the unconditional spec #298 measured, so #298's d5 row is an independent replication of the same effect. Fixed-effect inverse-variance pooling of the two 12-seed estimates (+0.0080, se 0.0033; +0.0062, se 0.0029):

pooled d5 effect: +0.0070, CI95 [+0.0027, +0.0112] — excludes zero.

The quick battery (d2-only) was not re-measured: with the gate off the spec is identical to master there, and today's d2 row confirms flat.

Guardrail compliance

• One codify/attack slot in this PR; no batching of independently-evidenced changes.
• composite_score formula and default batteries untouched (the legacy composite registry uses no gate keys, so its behavior is unchanged).
• Frozen presets untouched; the gate is opt-in per spec, following the --extra-highdim extend-don't-edit pattern.

Validation

• uv run pytest -q: 2014 passed, 1 skipped (coco not installed)
• uv run ruff format --check .: clean
• uv run ruff check on touched source files: clean (the 5 pre-existing tests/test_benchmark.py findings on master are unchanged)
• planning/SELF_IMPROVEMENT_LOG.md dated 2026-08-12 entry, TODO.md and planning/GOAL.md §5.1(d) updated in this PR

Follow-ups queued (in TODO.md)

• Re-measure the CMA-ES arm at d5 with this instrument — if it splits like NLSHADE_LBC, the same gate ships it.
• Codify-scan reading per_cell (GOAL §5.1(c)) so cell-conditional proposals become automatic.

🤖 Generated with Claude Code

https://claude.ai/code/session_01W7zJ73CD59Matb6HdSdrp2

---

Generated by Claude Code