v0.5.0

• feat(engine/brain): memory is in the loop — every glam run recalls relevant
  records from the project's local brain store (.glam/brain.db, hybrid retrieval,
  hard token cap, full provenance in the packed context) and writes a structured
  episode (task, outcome, decisions, files touched, models, real cost) back after
  the run. Header/receipt show what was recalled; --json gets a memory object.
  Controls: --no-memory, [memory] in glam.toml, GLAM_MEMORY=false. The
  self-contained npm bundle (no native store) says so honestly and runs without
  memory — it never fakes recall. Verified live: a teach run then a recall run
  against real Fireworks/GLM; a --no-memory control run could not answer. (#27)
• feat(scripts): adopt-claude-code.mjs — brownfield seed of glam adopt
  (#29): discovers existing Claude Code state on this machine (global + project
  CLAUDE.md, project memory, auto-memory) and imports it into the project brain,
  idempotently, so runs recall it from the first task. Claude Code files are
  read-only inputs — never modified.
• feat(cli): BEHAVIOR CHANGE — a run stopped by a budget/step/token ceiling now
  exits 3 instead of 0, so scripts and CI can tell a budget stop from done
  without parsing output. The documented, stable scheme (glam help /
  glam run --help): 0 done · 1 error · 2 usage error · 3 budget/step ceiling ·
  130 interrupted (128+SIGINT). (#23)
• fix(cli): the run header now shows the serving provider and model family
  (e.g. provider: fireworks model: deepseek-v4-flash (…)) instead of leaking the
  shared adapter's internal id (adapter: fireworks-glm) on DeepSeek runs; adapters
  now declare a stable provider id alongside their adapter id. (#24)
• fix(cli): the recorded to … ledger line and the monthly-budget alert passed
  the useColor function (always truthy) instead of the resolved boolean, forcing
  raw ANSI codes into piped glam run output. Found while verifying #23.