test(core): add real-Agent NFR-P7 idle-session memory bench

[bug-ops]

Summary

• Closes build a real-Agent-based memory bench to close the NFR-P7 idle-session measurement gap #5840. session persistence: fsync p99 exceeds NFR-P1 target, NFR-P7 idle-actor-memory unverified, replay reads whole log into memory #5445's standalone housing-cost harness could not construct a real Agent<LoopbackChannel> (private to zeph-core), so it only measured SessionActor::spawn's own housing (~65-93 KiB/actor). This adds an in-crate #[ignore]d test, nfr_p7_real_agent_idle_session_memory_floor, that spawns real SessionActors wrapping real Agent<LoopbackChannel> instances (one shared SkillRegistry across sessions via Agent::new_with_registry_arc, mirroring production's sharing pattern rather than a private registry per actor) and samples RSS the same way production's system_metrics module does, using the same batched/marginal methodology as the session persistence: fsync p99 exceeds NFR-P1 target, NFR-P7 idle-actor-memory unverified, replay reads whole log into memory #5445 harness.
• Measured combined (housing + Agent's owned state) floor: ~875-905 KiB/session — under NFR-P7's 1 MiB budget, but with far less headroom than the ~65-93 KiB housing-only figure implied. specs/068-session-persistence/nfr.md gains a rationale paragraph making explicit that this composite is a distinct, larger quantity than NFR-P7's own housing-only scope, rather than silently redefining it — and that this measurement is a synthetic lower bound (mock provider, one trivial skill, no real conversation history), not a production compliance proof.
• The test is #[ignore]d and matched by no CI workflow filter, so it never runs automatically; it's a manual tripwire, run via cargo nextest run -p zeph-core -E 'test(nfr_p7)' --run-ignored ignored-only --no-capture.

Review notes

• Team review (developer, tester, perf, security, impl-critic, code-reviewer) surfaced and resolved several issues before this PR: an initial per-actor-private-SkillRegistry construction was replaced with the shared-Arc pattern production actually uses; a doc comment mis-attributing the measured cost to the message-history buffer (this is a zero-conversation-turn floor, so history is empty) was corrected; the composite-vs-housing-only scope mismatch against NFR-P7's own budget was made explicit in both the test doc comment and the spec; a synthetic-lower-bound caveat and a decomposition TODO were added.
• A pre-existing, unrelated cargo deny advisory (RUSTSEC-2026-0204, via zeph-index→ignore→crossbeam-epoch) still fails the deny gate; this PR's diff does not touch that dependency chain.
• .local/testing/coverage-status.md and .local/testing/playbooks/session-persistence.md were updated in the main repository (not part of this diff, per project convention for shared testing docs).

Test plan

• cargo +nightly fmt --check
• cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings
• cargo nextest run --config-file .github/nextest.toml --workspace --features "desktop,ide,server,chat,pdf,scheduler" --lib --bins (12368 passed)
• RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace --features "desktop,ide,server,chat,pdf,scheduler"
• RUSTFLAGS="-D warnings" cargo check --workspace --all-targets --features desktop,ide,server,chat,pdf,scheduler --locked
• cargo test --doc -p zeph-core
• Manually ran the new bench: cargo nextest run -p zeph-core -E 'test(nfr_p7)' --run-ignored ignored-only --no-capture — passes, ~876-905 KiB/session across multiple runs