Unify ResourcePool + PressureSource traits — currently parallel, requires adapter shim

[joelteply]

airc-queue card

Coordinates work via the AIRC queue substrate (airc#562). Edit this card by commenting OR by running airc queue claim/airc queue release/airc queue heartbeat (later PRs).

{
  "kind": "airc-queue-card-v1",
  "id": "#1246",
  "owner": "claude-tab-2",
  "status": "merged",
  "evidence": "Adopted existing GitHub issue into airc queue.",
  "next_action": "Triage, claim, or close this adopted backlog card."
}

Close this issue when the work is done (status=merged/abandoned).

Original issue body

Pre-adoption body

Smell

Two parallel traits cover the same conceptual ground:

Trait	File	Methods	Origin
PressureSource	paging/broker.rs	name, pressure, evict_some, stats_snapshot	Phase 7 broker (PagedResourcePool-shaped)
ResourcePool	paging/pool.rs	tier_name, capacity_bytes, usage_bytes, evict_at_least, snapshot	#1228 (Docker / VRAM / NVMe tiers)

PagedResourcePool<K, V> happens to implement both via two separate manual impls. Tier pools that don't follow the per-key-page shape (DockerTierPool, future HF-cache tier, future system-RAM tier) only implement ResourcePool — and so couldn't register with the broker at all without the adapter shim added in #1245 (PR-4 of #1222).

Why it matters

Joel directive 2026-05-14: "code concurrency ONCE then incorporate it. Any hard coded into a subclass or at a lower level use of tokio etc are probably WRONG. ... Reimplenting concurrency is STUPID. ... I would call that out in issues as you find it."

Same principle applies to pressure/capacity surface. One canonical trait, every tier plugs in. The current situation forces an adapter type (ResourcePoolAdapter in paging/adapter.rs) that derives PressureSource from ResourcePool — works, but it's vestigial duplication and two impls drift without coordination.

Proposal sketch (NOT this PR)

1. Pick a winner. ResourcePool's surface is more general (capacity_bytes/usage_bytes is more diagnostic than pressure 0..1; evict_at_least(want_bytes) is more controllable than evict_some). PressureSource is just a derived-view over ResourcePool.
2. Move stats_snapshot() onto ResourcePool (it currently mixes with the diagnostic-only fields like hit_count that don't belong on a tier).
3. Delete PressureSource. PressureBroker takes Arc<dyn ResourcePool> directly.
4. Delete ResourcePoolAdapter (vestigial after the merge).
5. Update PagedResourcePool's manual ResourcePool impl to expose stats_snapshot natively (it already has the data internally).

Constraints

• Keep PagedResourcePool's existing pressure() / stats_blocking() public APIs working — those have other consumers (gpu/memory_manager.rs, possibly more). Add a fn pressure(&self) -> f64 { self.usage_bytes() as f64 / self.capacity_bytes() as f64 } default-method on ResourcePool so consumers reading pressure don't break.
• Migration is one PR per consumer (broker, then each ResourcePool impl, then deletion of PressureSource + adapter), to keep diffs reviewable.

Status

Filed by claude-tab-2 after #1245 (PR-4 of #1222) added the adapter shim that revealed the smell. Not blocking — the adapter is functional. Logged so the cleanup happens before more tier pools land and the adapter becomes architectural by accident.

Status log

• 2026-05-15T16:08Z — merged via PR refactor(paging,#1246): collapse PressureSource into ResourcePool — single trait, drop adapter shim #1264 @ 7b1cb01 (closed by github-actions[continuum#1142])