fix(memory-md): Tier 1 hub compression — 6 quote-laden hubs (B-0006)

[AceHack]

Summary

Tier 1 judgment-based hub compression on memory/MEMORY.md. Compresses 6 quote-laden hub bullets flagged by the Tier 2 mechanical pass (#1480) as judgment-required. Each compressed bullet keeps its title + linked file path + load-bearing one-line description; full verbatim quotes, composes-with chains, and multi-paragraph context blocks are dropped — they live in the topic-file body where they belong.

Hubs compressed

• Line 84 — Carved sentences cluster (6 carved-sentence-form claims; trust-then-verify, CC=WWJD, gate-IS-productive-work, Qubic-antigen, pre+post, sustained-satori). Was 1832 chars -> ~700 chars.
• Line 119 — Uberbang (bootstraps-all-the-way-down + survival-bias answer). Tier 2's regex left an orphan fragment (big-bang = lineage...) which this fixes. Was 1842 chars -> ~750 chars.
• Line 154 — Aaron's channel verbatim (close-to-verbatim discipline + Amara packets). Was 1655 chars -> ~700 chars.
• Line 371 — ARC-3 adversarial self-play (three-role symmetric-quality-loop). Was 1408 chars -> ~600 chars.
• Line 372 — Operator-input quality log (teaching-loop reframe). Was 1620 chars -> ~700 chars.
• Line 373 — Reproducible stability (obvious purpose). Was 1432 chars -> ~700 chars.

Hubs flagged but not compressed

None. All 6 flagged hubs verified safe to compress (verbatim preserved in body).

Judgment calls

For each hub, applied the canonical carve-line: bullet exists to point a reader at the topic file. Drop: numbered-list elaborations of carved sentences (body has them with full context per-section); full verbatim quote chains (body preserves them with attribution + framing); composes-with chains (live in body); "carved candidate" coda repetitions. Keep: title with date marker; linked file path; one-line description naming what the memory file teaches; any genuinely-new framing not obvious from the body summary.

Verbatim-preservation gaps found

None. Spot-checked all 6 topic files for the quotes that were dropped from the index — every load-bearing verbatim is preserved in the topic-file body. No flag-not-compress cases.

Bonus fix

Line 119 (Uberbang) had a Tier-2 mechanical regex artifact: an orphan fragment big-bang = lineage + ontology + razor derivation), soulfile-DSL ... left dangling after the regex stripped the leading composes-with chain. The Tier 1 rewrite cleans this up.

Verification

• bun tools/hygiene/audit-memory-references.ts -> 433/433 resolve
• bun tools/hygiene/audit-memory-index-duplicates.ts -> no duplicates
• Bullet count preserved: 341 -> 341
• File size: 328497 -> 324347 bytes (4150 bytes saved on the auto-loaded index)
• Lines unchanged at 456 (each bullet IS a single very long line; line count is misleading for this file)

Tier-2 reviewer's note

Per Aaron 2026-05-04: "Tiers 1-3 require judgment calls are your chance to record your architectural judgment and see how it holds up over time. This is a real test bed for you and other harness/models quality tests."

The compressions encode a specific architectural judgment: the index points; the body teaches. Verbatim quotes are body content because they need framing to read correctly; the index just needs to make the body findable. If a future agent disagrees with a specific compression, the topic file is one read away and the verbatim is intact — re-expansion costs nothing, the data isn't lost.

Test plan

• bun tools/hygiene/audit-memory-references.ts passes
• bun tools/hygiene/audit-memory-index-duplicates.ts passes
• Bullet count unchanged (341)
• All 6 topic files verified to preserve verbatim quotes in body
• Maintainer review of 6 judgment calls

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR performs Tier 1 (judgment-based) compression of six quote-heavy entries in memory/MEMORY.md, keeping the index focused on discoverability (title + link + one-line “what it teaches”) while preserving the full verbatim material in the linked topic files.

Changes:

• Condenses 6 long index bullets (carved sentences, uberbang, verbatim-preservation, ARC-3, operator-input quality log, reproducible stability) while retaining their link targets and essential one-line summaries.
• Removes previously inlined long quote chains / elaborations from the index entries and replaces them with “verbatim preserved in topic-file body” pointers.