Datastore query binary search + SPDX/CI hygiene

[facontidavide]

Two independent, low-risk fixes surfaced during a code review, based directly on main.

1 · perf(datastore): binary-search RangeCursor start and latestAt

ARCHITECTURE.md documents both as binary searches, but findFirstValid() and latestAt() linear-scanned the chunk deque and the rows within a chunk (O(C·R)). Committed chunks are never empty and are time-ordered (each chunk's t_min >= the previous chunk's t_max), so t_min/t_max are monotonic across the deque and timestamps are monotonic within a chunk — enabling std::lower_bound/std::upper_bound for an O(log C + log R) start position.

Behaviour-preserving: lower_bound keeps the first-duplicate range start; upper_bound keeps the last-duplicate / later-chunk-at-shared-boundary semantics of the previous reverse scans. Adds regression tests for duplicate timestamps, shared chunk boundaries, and single-point ranges. Makes the ARCHITECTURE.md "binary-searches" claim accurate.

2 · chore: SPDX headers + drop stale development CI trigger

number_parse.{hpp,cpp} (pj_base, Apache-2.0) were missing their SPDX-License-Identifier headers — the dual-license boundary depends on per-file headers, so add them. All four branch-triggered workflows still keyed on the deleted development branch; restrict them to main.

Verification

• Clean build under -Wall -Wextra -Werror + ASAN (./build.sh --debug).
• 62/62 tests pass (./test.sh), incl. the new query edge tests and derived_engine_test/engine_integration_test, which exercise latestAt through the MIMO join.
• All pre-commit hooks pass (clang-format v22 clean).

Release

PATCH-level — an internal behaviour-preserving optimization plus SPDX/CI hygiene. No API/ABI change.

🤖 Generated with Claude Code