feat(receipts): tool-execution receipts + epistemic-source classification (#251)

[tcconnally]

Summary

Implements #251: tool-execution receipts + epistemic-source classification (arXiv:2603.10060, NabaOS — "Tool Receipts, Not Zero-Knowledge Proofs"). Runtime-issued, HMAC-signed receipts per tool call (unforgeable by the model), a pramāṇa epistemic classifier, and a claim cross-check engine that flags fabricated executions, count misstatements, and false-absence claims. Extends the #235 evidence-level ladder and #224 external-artifact bindings with a receipt cross-check layer.

What changed

• ledger_agent/tool_receipts.py — build_tool_receipt/verify_tool_receipt (HMAC-SHA256 over id|tool_name|input_hash|output_hash|result_count|facts|timestamp, per the paper's field order; key resolution reuses the feat(receipts): evidence levels — separate signature authenticity from durable inclusion #235 key_registry path), a session-scoped ToolReceiptLedger with completeness (omitted-call) detection, pramāṇa classification (pratyakṣa/anumāna/upamāna/śabda/abhāva/ungrounded), six hallucination-type flags, five trust levels, and markdown epistemic-label rendering. receipt_to_evidence_hash() bridges tool receipts into evidence_hashes.
• benchmark/nyaya_verify_bench.py — deterministic 1,800-scenario NyayaVerifyBench adaptation (4 languages × 6 hallucination types × 50 + 150 clean) with per-type detection rates, false-positive rate, and verification-overhead measurement; fails the run below the issue's 90% / 20ms gates.
• docs/tool-receipts.md — receipt spec, pramāṇa table, 6-stage protocol, threat model, measured benchmark numbers vs the paper's 94.2% / <15ms, and the receipts-over-ZK cost argument (zkLLM ~180s/query vs receipts <20ms).
• tests/test_tool_receipts.py — signature/tamper/forgery coverage, all six hallucination types, completeness, trust levels, benchmark-generator determinism.

Success criteria (#251)

• Fabricated-tool-reference detection ≥ 90% on the adapted benchmark (paper: 94.2%) — measured 100.00% (0/600 false positives on clean claims; see PR comment / benchmark output).
• Verification overhead < 20 ms per response (paper: <15 ms) — measured 0.032 ms average / 0.032 ms median.
• Epistemic labels rendered on receipts and claim surfaces; absence claims verified against receipt sets.

Closes #251

Note on the benchmark numbers: the adapted NyayaVerifyBench is deterministic (seeded, programmatic ground truth rather than LLM-generated responses), so the 100% detection rate reflects the detector's completeness against exact injected fabrications rather than a claim of parity with the paper's 94.2% on their LLM-generated scenarios. The honest comparisons are: every fabricated reference is caught, false positives are zero, and overhead is ~0.03 ms vs the paper's <15 ms — receipts remain the cheapest practical verification layer for interactive agents.