feat(dig-node): live eviction test for inbound-demand tier + remove dead demand wrappers

[MichaelTaylor3d]

Closes dig_ecosystem#2013.

Context — rescoped closeout (loop-decider reconciled)

#2013 (wire the tier-1 inbound-demand signal into live relevance + eviction) is ~90% already delivered by epic #1934 PR-3 (eccc0cc):

• Eviction (b) — DONE: Node::module_tier (lib.rs:1352) = MAX(inbound_demand.tier(), tier-0 land ledger) → evict_modules_locked → plan_module_eviction sorts by evict_key=(tier.rank, mtime), so Tier0Precache evicts before Tier1Demand, LRU within tier.
• Relevance (a) — SUPERSEDED: post-#1934 the cache evicts by tier, not by relevance(); RelevanceInputs.local_read_count is consumed only by the tier-0 selector scoring speculative DHT-sampled candidates. Feeding demand there would be speculative wiring with ~zero behaviour gain + an Arc<InboundDemand> refactor — rejected per §2. Demand's keep-effect is delivered by the Tier1Demand tag.

So this PR is the honest closeout, not new wiring.

What changed

• Live eviction regression test — tests::inbound_demanded_module_survives_tier0_eviction_sweep (lib.rs): store A (note_inbound_demand → Tier1Demand, OLDER mtime) survives while store B (mark_tier0_land → Tier0Precache, NEWER mtime) is evicted, through the real module_tier→evict_modules_locked→plan_module_eviction path. Non-vacuous: A is the older file, so pure LRU-by-mtime would evict A and keep B — the opposite of the assertion; it can only pass because tier beats mtime. (The ticket's "B older" wording was self-contradictory — LRU evicts oldest first, making an older-B test vacuous — so A-older/B-newer is the correct construction.)
• Removed both #[allow(dead_code)] demand wrappers — deleted Node::inbound_demand_count + Node::inbound_demand_tier (redundant with module_tier's direct self.inbound_demand.tier()); repointed the 6 in-module test callers to self.inbound_demand.count/tier. Kept inbound_demand_entry_count (a real cache.stats consumer). Made tier0_live::mark_tier0_land pub(crate) for the new test (retains its non-test consumer). clippy -D warnings clean.
• Doc/SPEC drift corrected — inbound_demand.rs module doc now says inbound demand assigns the Tier1Demand tag giving eviction precedence (not "feeds relevance's local-demand term"); the stale SPEC "relevance demand term" claim was already removed by PR-3, so added a clarifying clause that local_read_count drives tier-0 candidate selection and is legitimately 0 for speculative candidates.

How verified

cargo fmt 0, clippy -p dig-node-core -D warnings clean, build green, cargo test -p dig-node-core --lib → 635 passed, 9 failed (all the known socket-bind sandbox tests). The new test + all inbound_demand tests pass.

Version

root [workspace.package].version 0.90.0 → 0.91.0; dig-node-core 0.39.0 → 0.40.0; Cargo.lock regenerated (dig-node-service=0.91.0, dig-node-core=0.40.0). Minor.

---

Generated by Claude Code