instructions-files-diet baseline — 2026-09-17 (operator-on-demand run; HEAD 096e6bda3)
A dated record. Re-derive; never cite these figures as authority.

§ 1 Render — bytes per consumer (baseline candidate composed from the committed rendition, no edits)
  gz content compose AGENTS.md --consumer root --candidate <copy of root.md>   exit 0
  root: total=48511 B  invariant=25006 B  compressible=354 B -> 0 B (presence attribution)  setpoint=lite
  baseline candidate byte-identical to AGENTS.md and to .gzkit/renditions/AGENTS.md/root.md

§ 2(a) Delivery — gz validate --instructions-files-budget   exit 0
  [surface-delivery-witness] AGENTS.md: 48511 B rendered against the codex delivery cap 65536 B — 17025 B of headroom
  [codex-delivery-witness]   AGENTS.md: 48511 B of 48511 B delivered by Codex — the whole surface reaches the agent
  Required delta against a vendor cap: 0 B. This run is driven by the operator's per-turn-load directive, not an overage.

§ 2(b) Delivery routing — AGENTS.md == root.md (cmp clean); no routing defect.

§ 2(c) Corpus budget (effective_corpus fold; 96 rows, 72 effective)
  tier invariant:     71 entries  25006 B   (verbatim floor, --rendition-floor-coherence)
  tier compressible:   1 entry      354 B   (governance-doctrine-surfaces; already absent verbatim from the rendition)
  generate_candidate (pure, nothing written): 31902 B
    = emitted entries 25360 + generated structure 537 + unowned carried forward 6005
    drops 16609 B of rendered text that no live corpus entry carries, inside sections declared corpus-owned
  non-corpus rendered text today: 48511 - 25360 - structure ~= 22.6 KB (unowned sections 6005 B of it; ratchet floor 6005)

Per-turn load, Claude Code (scratch script measure_load.py, read-only)
  48511  AGENTS.md
   2554  CLAUDE.md
   8550  .claude/rules/governance-core.md (no paths: frontmatter — loads always)
  17561  71 project skill descriptions (.claude/skills/*/SKILL.md name + description)
  77176  total repo-controlled per-turn
  (The operator's 34317 B figure for skill descriptions was not reproducible from repo files; the remainder is
   harness/plugin skills outside this repository.)

Session start
  21622  scripts/session_orientation.py stdout

On first edit under a path (nested AGENTS.md chain + path-scoped .claude/rules mirrors; same rule text delivered by both)
   73681  tests/**.py                 nested 29497 + mirrors 44184
  104623  src/gzkit/commands/**.py    nested 47686 + mirrors 56937
   78607  src/gzkit/governance/**.py  nested 34698 + mirrors 43909
   70335  docs/design/adr/**/obpis/*  nested 35135 + mirrors 35200
   72152  .gzkit/skills/**            nested 36055 + mirrors 36097
   58756  .gzkit/rules/**             nested 29415 + mirrors 29341
   41506  .gzkit/chores/**            nested 20844 + mirrors 20662
   14465  docs/governance/**          nested  7330 + mirrors  7135

chores advise (2026-09-17): criteria 1-5 PASS; criterion 6 `uv run -m unittest -q` reported FAIL = "Timed out after 120s"
  (the advise runner's ceiling, not a test failure; the suite's own result was not observed in that run).
