Track remaining query tier-boundary correctness and corpus follow-ups #24116

Track the engine and corpus work intentionally left outside the surgical iteration of #24097 and #24098. These are explicit follow-ups, not a claim that current query output or the full corpus is correct. The accepted scope repairs NI-1, strengthens independently keyed contracts, rules a coarse seam envelope and narrows conventional-tail coverage to its actual 4-second grid.

The unaligned-ring regression R2 has its own issue: #24115.

Correctness work

• F9: coarse seam anchoring. In the shifted CASE-040 fixture, neighboring coarse averages are 200 and 10400/43; the approved estimate must stay between them. Master and forced tier 1 return approximately 210.465; automatic Fix query coverage and handoffs across storage-tier retention #24097 returns 650. The revised shifted-constant-seam contract remains red. Evaluate coarse-average anchoring without changing ordinary single-tier interpolation or silently replacing fixture-derived bounds with observed output.
• F6/P-04: incomplete first coarse records. Preserve fine samples omitted from the first post-enable coarse record. The strengthened partial-first-record-{sum,average,min,max,metadata} fixture puts 2000 at offset 620 in samples 601..660: SUM39210, AVG653.5, MIN601, MAX2000, anomaly25%. Forced tier 0 is the truth control. All five automatic claims remain independently keyed.
• Raw-fine fallback at page-hole boundaries. Retained fine samples 661..690 and 1261..1290 should contribute constant7 to rows720 and1320. Master and the current head return NULL after selecting a coarse tier with a hole. late-tier-boundary-average records these failures separately from late-tier-gap-null.
• P-01: conventional live-tail rows wider than4s. The current conventional-tail contracts read both tiers on aligned4s rows only. Supplied8s probe gives AVG11 instead of11.5, MAX11 instead of12 and anomaly0% instead of50%; SUM is correct for that probe. Add separate wider-grid coverage with exact fixture truth before an engine remedy.
• P-02/F7: isolated coarse tail. The second isolated-tail row(+1200,+1400] remains NULL instead of7 on master/current head. isolated-coarse-tail already exposes it; investigate exhausted-plan handling.
• R1: first-row island skip. Row-top plan expiry can skip a coarse island before its first record is read for AVERAGE/MIN/MAX. Add a fixture whose first output row contains the island end and preserve both island and later fine tail.
• O1: small-grouping coarse-to-fine overlap. With grouping4, plans ending at124/128/130 and17s rows, the tail lookbehind re-reads fine samples123/124. Supplied SUM2321 versus fixture2074. This was present before the NI-1 update. Cover groupings<=5 and retain stock-grouping behavior.
• P-11: numeric future-record loss after a page hole. A single-tier first point(2000,2001] can be consumed in the previous output row, yielding MIN2002 instead of2001 in(2000,2005] and omitting its statistics. The coarser-only NI-1 spacer fix does not solve this general path. A validity flag alone is not sufficient proof; review row-boundary consumption.
• P-12/P-13/P-07: RAM/ALLOC connected-boundary geometry. A fine ring starting inside a coarse record can supply empty placeholders that discard the coarse record; a slot start equal to an island end can cause SUM loss. Test connected overlap, equal endpoints and the slot-start retention convention independently of the disconnected clamp.

Diagnostics, performance and harness decisions

• F12: debug plans. The query_plan diagnostic still lists head plans whose storage handles never initialized, although physical-work counters are now accurate. Decide how to represent proposed versus executed plans without changing counter semantics.
• F15: resampling scale. Establish and test the intended time_resampling behavior when automatic retention recovery brings a finer head into a selected coarse query. Do not adopt a numerical expectation before the contract is ruled.
• F11: remaining seam-bookkeeping cost. Measure the per-point arithmetic on representative multi-tier workloads. The previously stated2–4% CPU figure has not been independently established; no performance regression magnitude is claimed here.
• F17: clipped singleton preference. A selected tier clipped to one timestamp supplies no interval boundary, so synthetic geometry can apply older-side preference to a newer region. No production trigger has been demonstrated. Determine reach and contract relevance before changing the approved tier order.
• TESTS-7: shared setup classification. Registration before setup currently reports explicitly incomplete/NOT RUN contracts and the full suite fails. Decide whether setup failure should also classify affected keys as BROKEN. Preserve clean skips, incomplete components, sticky failure and independence between sibling contracts; avoid marking every key broken because one assertion fails.
• P-10/P-16: harness evidence and shutdown. Preserve logs across restart; make forced shutdown distinguishable from orderly success; account for settle-timeout flakes and source-built UBSan reports on stderr. Check shutdown behavior using harness-owned process identities only.
• P-15: datatable timezone contract. Absolute failure rosters currently require TZ=UTC. Expected date fields are zero-padded while the engine strips leading zeros, causing an additional deterministic failure in Europe/Athens. Rule and fix the formatter expectation independently.

P-14 ruling: SES/DES carrying smoothing state across gaps is existing design behavior. Document the scope when adding gap assertions; do not impose NULL on those groupings merely because ordinary AVERAGE gaps are NULL.

Evidence surfaces: production query planner/executor, independently keyed CASE-040 fixtures and manifests. The listed comparative probe results are preserved investigation evidence; each implementation requires its own failing automated reproducer and a recorded design decision where semantics are unresolved.

Close individual items by linking the fixing PR, an evidence-backed rejection, or a narrower owning issue. Preserve separate invariant keys and compare full same-corpus failure rosters.