[Levalicious]

Problem

The server never validated call arguments against the advertised inputSchemas, so undeclared keys were silently dropped. Concretely: open_nodes called with cursor (the get_neighbors page key) instead of relationCursor re-served page 1 with an identical nextCursor — every relation past page 1 was unreachable, and it looked like open_nodes ignored cursors. An off-enum direction: "both" also ran without complaint.

Fix

• The tool list is bound to toolDefinitions; ListTools returns it and dispatch validates every call against the same schemas (validateToolArgs).
• Undeclared keys and off-enum values → VALIDATION_FAILED, every offender listed in details (argument, message, allowed), tool does not run (R2 validate-first).
• A misnamed cursor key names the cursor parameter(s) the tool actually takes.
• docs/api-error-policy.md: R4 now explicitly covers arguments.

Existing per-field checks (LIMIT_EXCEEDED etc.) are untouched.

Tests

tests/arg-validation.test.ts (6): 5 fail on develop, all pass here; the 6th drains a 120-edge backward list via relationCursor (guards the correct path). Full suite 202/202, fuzz protoErr = 0.

🤖 Generated with Claude Code

https://claude.ai/code/session_01B7UQfn7Vue83jAPFaWfmAZ

[github-actions]

Per-op benchmark — p50 cycles (base vs head)

base runs: 5, head runs: 5 (median of mins) · freq-correction ×1.0038 · threshold head−base > max(40 cyc, 5%)

op	base	head*	Δ	Δ%
lookup	75.2	73	-2.2	-2.9%
read_entity	36.5	37	+0.5	+1.4%
neighbors_d1	584.4	587	+2.6	+0.4%
neighbors_d2	2025.1	1999	-26.1	-1.3%
find_path_d6	163514	159914	-3600.0	-2.2%
search	969670	1002076	+32406.0	+3.3%
entities_by_type	14404	14485	+81.0	+0.6%
random_walk_d5	1207.1	1193	-14.1	-1.2%
inc_walker_visit	15.6	16	+0.4	+2.6%
structural_sample	2315430	2329747	+14317.0	+0.6%
compute_merw_psi	686088	709367	+23279.0	+3.4%
create_entity	884	887	+3.0	+0.3%
create_relation	780	783	+3.0	+0.4%

* head frequency-corrected to base via the reference op.

No regressions beyond threshold.