feat(retrieval): use_type_aware_compression flips to default-on (#769) [stacked on #879]

[robotrocketscience]

Closes #769.

Stacked on #879 (which closes #878 / compose-reconciliation). Set base to feat/issue-878-compress-cluster-compose until #879 lands, then this rebases to main.

Summary

Flips resolve_use_type_aware_compression() default False → True. The bench evidence cleared empirically against the post-#815 lab-side compression_a* corpora — see #769 bench-results comment:

Gate	Mean	Per-row positives	Per-row regressions	Spec threshold
A2 recall@k uplift	+0.3267	15 / 25	0 / 25	strict > 0 aggregate, per-class non-regression rider
A4 fidelity delta	+0.0085	2 / 15 (rows with headroom)	0 / 15	≥ −0.005 band

Per-class non-regression rider (drop / summarise / verbatim cohorts each non-regressing) holds; all-fact cohort delta is exactly 0.0000 per the spec's stated degenerate case.

Why this is the right shape

resolve_use_type_aware_compression() docstring contract: "the bench gate flips the default after lab-side benchmark evidence clears." That precondition is now met. Compose-compatibility with the (already default-on) use_intentional_clustering flag — the structural prerequisite that almost blocked an earlier flip-attempt mid-session — ships in #879 / #878 right before this.

Mechanism

• src/aelfrice/retrieval.py:1726 — return False → return True. Docstring's "Default: False" → "Default: True" with bench-evidence callout + feat(retrieval): use_intentional_clustering × use_type_aware_compression compose-reconciliation (blocks #769) #878 compose-link.
• Precedence (env > kwarg > TOML > default) is unchanged; only the default value flips. Operators on v2.x parity:
  • [retrieval] use_type_aware_compression = false in .aelfrice.toml, or
  • AELFRICE_TYPE_AWARE_COMPRESSION=0 env var.

Tests

Mirrors the #436 default-flip precedent (ab4462d7):

• test_default_is_off → test_default_is_on
• test_pack_byte_identical_when_flag_off → test_pack_byte_identical_when_flag_on — ON-byte-identity invariant supersedes the prior OFF-byte-identity invariant; the assertion that default == explicit is preserved, only the explicit-arm flips.
• test_retrieve_pack_byte_identical_when_flag_off → test_retrieve_pack_byte_identical_when_flag_on — same shape on the bare retrieve() path.
• test_retrieve_env_var_enables_compression → test_retrieve_env_var_disables_compression — the env-var path now tests rollback to OFF (env="0") rather than activation (env="1"), matching the post-flip precedence semantics. The bench-harness path that originally motivated this test still works — AELFRICE_TYPE_AWARE_COMPRESSION=1 is the no-op default now.
• test_default_call_leaves_compressed_empty → test_default_call_populates_compressed — assertion flipped to match the new default.
• test_env_garbage_falls_through — default-arm assertion flipped.

Full suite green locally on stacked HEAD (4656 passed, 64 skipped, 75 xfailed).

Doc surfaces

Three places still asserted v2.1-era default-OFF posture — all corrected in a separate atomic commit:

• docs/design/feature-type-aware-compression.md status banner.
• docs/user/CONFIG.md — TOML example flips to true, flag-doc section reframes default + precedence narrative, tier-table summary line.
• docs/concepts/ROADMAP.md v3.0 row — A2 bullet amended to note A4 also cleared and the flip landed in Flip default: use_type_aware_compression on by default — gated on A2 + A4 bench run #769.

Determinism (#605)

Precedence chain unchanged; only the default value flipped. Replay with same inputs reproduces same output (now compression-ON unless explicit override). The compression render itself is pure over (belief content, lock state, retention class) — no wall-clock, no random sampling.

Discretion

Diff against main is clean against the pre-push deny-list.

What this closes

Three R&D rounds (R7–R18 in 2026-05-14/15) + the @Ilya0527 statistical-power exchange (2026-05-18) + the 2026-05-20 operator verdict to flip — see #769 thread. The campaign's verdict-without-numbers state finally collides with numbers: A2 cleared, A4 cleared, no regressions, compose-reconciliation in place.

[sourcery-ai]

Sorry @robotrocketscience, you have reached your weekly rate limit of 500000 diff characters.

Please try again later or upgrade to continue using Sourcery

[coderabbitai]

Warning

Rate limit exceeded

@robotrocketscience has exceeded the limit for the number of commits that can be reviewed per hour. Please wait 32 minutes and 6 seconds before requesting another review.

You’ve run out of usage credits. Purchase more in the billing tab.

⌛ How to resolve this issue?

After the wait time has elapsed, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

We recommend that you space out your commits to avoid hitting the rate limit.

🚦 How do rate limits work?

CodeRabbit enforces hourly rate limits for each developer per organization.

Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout.

Please see our FAQ for further information.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 6082660e-0f5b-4750-aed7-5340f68f014a

📥 Commits

Reviewing files that changed from the base of the PR and between a1f2c30 and 884ca5a.

📒 Files selected for processing (7)

• CHANGELOG/v3.md
• docs/concepts/ROADMAP.md
• docs/design/feature-type-aware-compression.md
• docs/user/CONFIG.md
• src/aelfrice/retrieval.py
• tests/test_compression_integration.py
• tests/test_context_rebuilder.py

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feat/issue-769-flip-compression-default

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[robotrocketscience]

Stacked on #879 — do not review/label-merge this PR until #879 lands. Once #879 merges, I (or any session) will rebase this branch on main and add attn:review then.

[github-actions]

PR-size soft cap

This PR is over the advisory size threshold:

• 131 changed lines (limit: 200)
• 7 changed files (limit: 3)

Bigger PRs collide with more open work, which under the parallel-session workflow tends to produce repeated attn:merge-conflict cycles (see #602). When practical, split into smaller PRs that each touch a focused surface.

This is advisory only — nothing is blocked. If the size is intentional (large refactor, module removal, generated code), apply the size:override label and this comment will be removed on the next push.

[github-actions]

merge-train: merged 884ca5a → main via FF push.