D24: admit build_tdigest_where to the digest family; give composition an honest composability declaration

[espg]

🤖 from Claude

Filed at espg's direction (in-session 2026-08-24, ruling: admit). Split from PR #510's standing question (1) — the gate drift its phase-1 characterization surfaced and deliberately did not unify.

The drift, and the ruling

build_tdigest_where (the strata builder) is admitted by the spill gate (_TDIGEST_SPILL_FUNCTIONS, issue #370: stratum membership is per-photon row selection before the build, so per-block strata digests merge like any digest) but not by D24's composability gate — so strata fields classify {class: none}. This is not hypothetical: the published CA store's manifest declares h_tdigest_signal/h_tdigest_noise/composition all class: none (verified live on atl03_tdigest_o9.zarr/morton_hive.json), while every one of its 2,726 shards was built by k-way merging per-block strata partials under spill — the store's bytes already rest on the law D24 declines to state. Espg ruled: admit.

Scope (two admissions, one honest declaration)

1. build_tdigest_where → the digest-family registry (PR D24 composability: admit build_waveform_digest via a shared digest-family registry (issue #508) #510's _DIGEST_FAMILY_FUNCTIONS): same k-way law, same (2,) centroid element, delegates to build_tdigest — pure registry admission plus parity tests (strata payload + located/temporal companions through the fold; the WHERE clause has no effect on merge semantics).
2. composition: merges by a different, exact law — merge_composition_kway over packed (word, n_signal) pairs (issues Carry ATL03 signal confidence through aggregation: signal/noise split + t-digest encoding for signal-weighted photon flux #321/spill: cross-block merge for located / strata / composition fields (lift the single-block restriction) #370; also already spill-admitted). D24's approximate arm is digest-shaped, so composition needs its own honest class/method declaration rather than shoehorning — determine the right shape (likely an exact-law entry with its packed-ragged encoding declared); if that needs a new D24 method/class token, the spec page and conformance fixtures move in the same PR (Promote zagg-ragged/1 + digest + composition byte layouts to a normative spec (reader-migration gate) #340 rule). Without this, a strata pyramid loses its signal-fraction companion at overview levels — arguably the most useful browse quantity.

Why now

The pyramid retrofit sweep planned for the 0.50 fleet declares fields from the classifier — admitting strata (+composition) is what makes the CA ATL03 retrofit produce digest pyramids instead of count-only. Retrofit mechanics note: leaf columns were never written for these fields (class none at build time), so the retrofit uses the leaf-reading path (declare_pyramid + cascade-from-leaves) — more reads, zero re-aggregation.

Refs: #508 / PR #510 (the registry + the raised drift, issue comment 5398201934), #370, #321, #334/#337 (strata design), #340 (spec+fixtures rule).

[espg]

🤖 from Claude

Implementation plan (espg-directed: separate PR, not a #510 phase — the strata admission is mechanical but composition needs an honest declaration of its own, and #510 stays a finished reviewable unit).

Phase 1 — characterization. Pin today's behavior for both fields: D24 class (none for both), the spill gate's admissions, and the CA-manifest shape as the known-answer fixture (h_tdigest_signal/h_tdigest_noise/composition all none, count exact — the live store's declaration, so the fix's before/after is pinned against reality).

Phase 2 — build_tdigest_where admission. Add it to _DIGEST_FAMILY_FUNCTIONS; k-way parity tests for the strata payload AND both companions (located words to common ancestor, temporal words to the toc envelope) through the fold; template-classification test: the shipped strata template + pyramid = {} declares {class: approximate, method: <tdigest law>, delta, overview_delta, temporal/located as declared} for both strata fields.

Phase 3 — composition's declaration. Determine the honest D24 shape for merge_composition_kway over packed (word, n_signal) pairs — an exact merge law with a non-scalar encoding, so neither existing arm fits as-is. Options to weigh in-PR (decision recorded in the body, spec §/fixtures updated in the same PR if a new class/method token is needed, per the #340 rule): extend the exact arm with the packed method, or a third class token. Fold parity: per-level composition equals the k-way merge of contributors, and the signal-fraction derived from it matches the leaf truth.

Phase 4 — end to end. Synthetic strata store: template → build → overview fold; per-level weight_total parity for both strata; composition companion present and correct at every level; the CA-manifest before/after known-answer test.

Deployment note (recorded, not in this PR): classification is template-time worker-side — takes effect on the 0.50 fleet; the CA retrofit uses the leaf-reading path (declare_pyramid + cascade) since strata columns were never written.

Branch claude/515-strata-composability from main after PR #510 merges (it provides the registry). Phased draft PR per §2, opus review + fold per phase.