CLAUDE.md (= AGENTS.md) is auto-loaded in full at the start of every
session. It had grown to 398 lines / 170 KB (~78k tokens, 9% of a 1M
context) because every completed phase appended its full record to it.
Almost all of that is history: valuable, but needed on demand, not on
every turn.

Move lines 66-397 verbatim into docs/progress/, split by phase, and
keep in CLAUDE.md only what every session needs: positioning, the
7-layer architecture, the numbered design invariants (referenced by
name from source docstrings), the correctness red lines, environments,
the secret discipline, dev conventions, the current status snapshot,
the standing authorizations, the working protocol, the methodology
rules, an index of the archive, and the disclosures/roadmap that are
currently in force.

Nothing is rewritten or summarised away. A reconstruction script
verified that concatenating the archive bodies reproduces the original
332 lines exactly, line for line, and the PR ledger line likewise.
Lines 1-59 and 61-64 of CLAUDE.md are byte-identical to before; only
the Git bullet was replaced by a shorter one pointing at
docs/progress/pr_ledger.md.

CLAUDE.md: 398 -> 151 lines, 170316 -> 14469 bytes.

To keep it from growing back, the archive section states the rule
explicitly: new phase records go to docs/progress/, and CLAUDE.md
maintains only three moving parts (status table, disclosures,
roadmap).

Gates: pytest 2537 passed, ruff clean, CLAUDE.md == AGENTS.md.