chore(release): 0.16.4 — ADR-037 Slice B (wire-evidence echo on /gate) - #96

* feat(sdk): wire-evidence echo capture from /gate response (ADR-037 Slice B)

Implements the SDK-side of the protocol-3→4 additive bump. The
backend's /gate response now echoes the SDK-supplied action_digest
and a policy_hash slot (None today; Slice D wires per-request
computation). Pre-v4 backends omit both keys entirely via
skip_serializing_if = "Option::is_none", so the v4 SDK degrades
cleanly to None on a v3 backend — no false positive.

Source-side wiring:

1. NULLRUN_PROTOCOL_VERSION = 4 (src/nullrun/transport.py). The
   X-NULLRUN-PROTOCOL header now serialises the bumped value via
   the single source of truth; tests pin to str(NULLRUN_PROTOCOL_VERSION)
   so a future bump doesn't sweep this file again.
   NullRunProtocolError.user_action and docs/errors/NR-P001.md
   updated to point operators at X-NULLRUN-PROTOCOL: 4.

2. Two new contextvars in src/nullrun/context.py —
   _last_gate_action_digest_var, _last_gate_policy_hash_var — with
   public accessors get_last_gate_action_digest() /
   get_last_gate_policy_hash() and setters set_last_gate_action_digest() /
   set_last_gate_policy_hash(). clear_server_minted_execution_id
   also drops the v4 slots so a /check in one block never leaks
   a stale echo into a /track in a sibling block.

3. runtime._capture_wire_evidence (extracted from
   _capture_server_minted_execution_id so the two captures share
   a call site but have distinct log lines). Reads action_digest
   + policy_hash off the response, defensively validates type
   (drops non-str at WARNING), and stores into the contextvars.
   Called from _capture_server_minted_execution_id on the same
   /check lifetime so execution_id and action_digest always refer
   to the same gate decision.

4. ServerCapabilities.wire_evidence_echo (src/nullrun/capabilities.py) —
   informational flag surfaced by /api/v1/capabilities. NOT
   included in is_v3_ready() (informational, not a hard gate).
   Defaults to False on pre-v4 backends.

5. tests/test_slice_b_wire_evidence.py (10 new tests). Pins the
   SDK-side of the v3→4 additive bump: protocol-constant value,
   header serialisation, capture from /gate response (happy path,
   policy_hash-when-present, both-set), tolerance of pre-v4
   backends, tolerance of malformed wire values, tolerance of
   non-dict response, clear semantics, source-of-truth wiring.

6. tests/test_capabilities.py — 2 new assertions:
   test_parse_capabilities_wire_evidence_echo_v4_backend and
   test_parse_capabilities_v4_protocol_range.

7. tests/contract/test_audit_wire.py + tests/test_v3_wire_contract.py
   — header assertions now source str(NULLRUN_PROTOCOL_VERSION)
   instead of the literal "3" so a future bump doesn't require
   sweeping either file. Class names kept for git-blame continuity.

8. README.md — alpha-status line + roadmap table updated:
   v0.15 → v0.15.x, v0.16 → v0.16.x with the new highlights,
   v0.17 for OpenTelemetry exporter / Redis-backed offline queue /
   hardened init contract that previously sat under v0.16.

* docs: NR-P001 user_action bumps to protocol 4 reference

NullRunProtocolError.user_action and docs/errors/NR-P001.md
updated to point operators at X-NULLRUN-PROTOCOL: 4 (was 3).
The v3→v4 bump is additive: min_protocol_version stays at 2,
so v3 SDKs continue to work against a v4 backend. The new
user_action tells operators running a pre-v4 SDK against a
v4+ backend to upgrade.

* chore(release): 0.16.4 — version bump + CHANGELOG

- pyproject.toml + src/nullrun/__version__.py: 0.16.3 → 0.16.4.
- CHANGELOG.md: insert [0.16.4] - 2026-08-31 at the top with full
  descriptions of the Slice B wire-evidence echo (capture helper,
  contextvars, capabilities flag, test pins), the protocol version
  bump, the wire-format additive nature, and the 1613/4 pytest
  + ruff-clean + mypy-clean verification status.

Verified: pytest 1613 passed / 4 skipped (12 more than 0.16.3,
accounting for the 10 new Slice B pins + 2 new capabilities
assertions); ruff check src tests all checks pass; mypy src/nullrun
no issues reported in 37 source files. No new wire fields beyond
the two echoed slots — both fields echo already-computed values,
no new hashing/computation introduced on either side.

Refs: ADR-037 Slice B (2026-08-31); backend commit 8dbeaf4d
(parity CI test analogue for SDK side); docs/errors/NR-P001.md.