feat(api): error envelope + validate-first atomicity (policy R1/R2/R4)

[Levalicious]

Implements the error half of docs/api-error-policy.md (adopted 2026-10-08): the server no longer hides failures from the model.

What changed

• R1 — visible failures. ToolError + stable codes (ENTITY_NOT_FOUND, ENDPOINT_MISSING, COLLISION, LIMIT_EXCEEDED, INVALID_REGEX, INVALID_FILE, NO_MATCHES, TYPE_NOT_FOUND, VALIDATION_FAILED). The request handler converts them into isError: true with a JSON envelope {error: {tool, code, message, details}}. Protocol errors are now reserved for genuine server faults.
• R2 — validate-first, atomic. create_entities (collisions were checked mid-write), create_relations, add_observations (validated inside the write loop → partial batches), sequentialthinking — all validate the whole request before touching state; failures list every offender in one round-trip.
• R4 — drops→errors. Missing relation endpoints, all-missing open_nodes, unknown get_entities_by_type types, unknown get_neighbors/random_walk starts, invalid regex, kb_load file errors, unlinked thoughts. Duplicates and ensure-absent deletes keep succeeding (their ledgers arrive in the follow-up ledger PR).
• search_nodes's NL-guard now returns the same envelope shape (NO_MATCHES + suggestedRegex detail).

Evidence

• 13 existing protocol-error assertions converted to envelope checks (parseToolError helper in test-utils); new API error policy test block proves atomicity (failed batches leave nothing behind) and envelope shape.
• Fuzz (30 agents, ~366k calls): protoErr = 0 across every tool — all 25.8k create_entities rejections, 14.3k relation rejections, 10.3k get_neighbors rejections etc. now arrive as visible toolErr.
• Full suite: 183/183.

🤖 Generated with Claude Code

[github-actions]

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

base runs: 5, head runs: 5 (median of mins) · freq-correction ×1.0025 · threshold head−base > max(40 cyc, 5%)

op	base	head*	Δ	Δ%
lookup	115.1	116	+0.9	+0.8%
read_entity	40.1	40	-0.1	-0.2%
neighbors_d1	797.1	803	+5.9	+0.7%
neighbors_d2	2593.2	2630	+36.8	+1.4%
find_path_d6	177552	178534	+982.0	+0.6%
search	1172840	1175146	+2306.0	+0.2%
entities_by_type	14039	14073	+34.0	+0.2%
random_walk_d5	1321.6	1309	-12.6	-1.0%
inc_walker_visit	17.5	18	+0.5	+2.9%
structural_sample	2576249	2587993	+11744.0	+0.5%
compute_merw_psi	905814	910844	+5030.0	+0.6%
create_entity	833	810	-23.0	-2.8%
create_relation	784	786	+2.0	+0.3%

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

No regressions beyond threshold.