fix: v0.4.1 — RELAY ecosystem audit fixes, pass 2

[SoundMatt]

Summary

Second round of fixes from the SoundMatt x-Net ecosystem audit against cpp-LIN. Applies 15 of the 15 diffs marked "ready" in the audit register (cpp-LIN-N2-01, cpp-LIN-01/02/03/05/06/07/11/12/13/14/15/16/17/18), reviewed and cross-checked individually against the underlying findings before applying. Bumps to v0.4.1.

• Add verify_checksum() to actually validate a received checksum against the locally-computed value (ISO 17987 frame reception). HARA SG-03 downgraded from "Implemented" to "Partial" — no real (non-virtual) bus backend calls the new function yet, and no dedicated unit test was included in the source diff, so the safety claim is intentionally left honest rather than "closed."
• Replace the generic ErrTimeout returned by send_header on no slave response with a distinct lin::Errc::no_response sentinel that still compares equivalent to a timeout for existing callers. (The pre-existing dead lin::ErrNoResponse exception class and its false-coverage test are untouched — that cleanup is bundled with the still-open advisory cpp-LIN-08 test-gap item, out of scope here.)
• from_message now rejects an unrecognized lin.checksum_type instead of silently defaulting to Classic.
• Bump declared RELAY spec version 1.11 → 2.0 across headers, CLI JSON, README, HARA, and requirements (RELAY v2.0's breaking change is scoped to RCP and does not alter LIN's canonical types). Also fixed a stale v1.11 reference in HARA.md:134 that neither source diff touched.
• Fix testdata/relay-vectors/lin-frame.json to encode checksum_type as "enhanced" instead of the bare numeral "1" — independently re-derived: vector checksum 73 matches the Enhanced algorithm for id=16, payload {0xAA,0xBB}, not Classic (153).
• master::Node::set_schedule now accepts an empty schedule (spec §8.3), matching the virtual bus.
• DropOldest back-pressure now counts an eviction as a drop instead of a delivery.
• Frame::checksum_type now defaults to Classic (the enum zero value), matching the other language bindings.
• capabilities no longer lists "virtual" in both features and transports.
• E2E Receiver::unwrap no longer advances its sequence counter on a rejected frame, so a persistent shifted stream keeps being flagged instead of silently resyncing after one error.
• Define the previously-dangling REQ-CLI-001..006 / REQ-SEC-013 requirements referenced by existing fusa:req tags; correct three requirement rationales that cited the wrong Safety Goal; fix HARA's error-handling citation (§15 → §5).
• Add CHANGELOG.md.

Not included (advisory only, no diff provided — still open)

• cpp-LIN-04: relay.hpp depends backwards on lin::Chan — needs an invasive namespace refactor (move Chan<T> into namespace relay); design decision, not attempted.
• cpp-LIN-08: Tests hardcode inline JSON instead of loading testdata/relay-vectors/, and mirror the (now-fixed) ErrNoResponse bug instead of asserting the real sentinel — multi-concern test-suite rework, not attempted.
• cpp-LIN-09: .fusa-reqs.json (145 reqs after this PR) vs requirements/requirements.json (100 reqs) disagree on population — canonical-file governance decision needed.
• cpp-LIN-10: Structural frame-ID errors mapped to ErrPayloadTooLarge — contested sentinel mapping, needs a documented adapter policy decision.

Test plan

• cmake -S . -B build -DCMAKE_BUILD_TYPE=Debug configures cleanly
• cmake --build build builds cleanly, no new warnings
• ctest --test-dir build --output-on-failure — 171/171 pass (reproduced locally, not just trusting the audit register's self-report)
• Spot-checked cpp-lin-cli version / capabilities output at runtime — reports 0.4.1 / spec_version: 2.0 / de-duplicated features