v4.5.0 "Reach"

[4.5.0] — 2026-06-21 — "Reach"

Why: two ways to put the right context in front of an agent with less
waste — displace redundant context across providers instead of concatenating it,
and point a spawned sub-agent at the files most likely related to the one it is
working on. Both are structural and never-worse by construction; neither is a
cost/bill claim.

Added

• Sub-agent focal reach (#139) — when a sub-agent starts, the context broker
  now appends the files most likely related to the parent's most-recently-read
  file: graph-adjacent files first (real call/import edges), then path-based
  reach (same-directory siblings, test↔implementation counterparts). Tiered so a
  real graph hit is never displaced by a weaker path guess; de-duplicated;
  capped. Lets a sub-agent skip the grep-around discovery step. Degrades silently
  to the existing top-entities map when no focal file can be inferred.
• Reach benchmarks — bench/recall-coverage gains a path-reach mode
  (ENGRAM_BENCH_PATHREACH=1); new bench/cochange-holdout is a non-circular
  temporal-holdout test (learn co-change from past commits, predict held-out
  future co-change) against a popularity baseline. Honest result in the output:
  co-change alone is weak, but combined with path-reach it beats the baseline.

Changed

• Cross-provider displacement — the resolver now de-duplicates and
  blend-ranks context across providers and discards the redundancy, instead of
  concatenating each provider's section. The Stop summary reports ~N redundant tokens displaced across providers (structural — not a bill saving).
• Never-worse gate on the enriched Read packet — an intercepted Read only
  serves the enriched (structure + provider) packet when it is strictly smaller
  than the raw file; otherwise engram falls back to the graph-only summary
  (already proven smaller). Guarantees an intercepted Read never injects more
  tokens than the file it replaces.