Measure TIN strategy, visibility, and CTID layout behavior

[bweis]

Small TIN fixtures hid important differences in filtering, prepared plans, visibility, and index layout. This adds bounded remote experiments through the existing benchmarks/tin.py experiment command and records 4,896 query probes up to one million Wikipedia articles, plus 24 concurrency windows. The larger deployment was a user-confirmed PS-160 ARM with EBS, 2 vCPU and 16 GiB RAM.

The results support filtered TID pushdown as a concrete Stannum optimization target: it beat TIN's automatic choice on both machines and at one million rows, and the result held after controlling visibility. They also show costly forced-generic dynamic-LIMIT plans, count timings dominated by visibility state, and compact physical-TID indexing across a 33-fold change in heap-page span. Checked-in summaries and 21 complete plan examples distinguish observations from implementation hypotheses.

The collector retains SQL, complete plans, source snapshots, segment state and client samples; uses session-only controls and uniquely owned schemas; and records individual query errors without hiding incomplete runs. Credentials are excluded. A million-row REINDEX and cleanup initially hit a three-second lock timeout; a separately recorded recovery completed REINDEX, passed TIN fsck, and verified cleanup. A live lock-holder regression verified the bounded maintenance wait and setting restoration.

Validation:

• 118 Python harness tests and source-attribution checks pass locally.
• Forced ranking score-sequence checks, multi-index membership/count comparisons, and cross-layout count/score checks passed.
• Four recorded query errors came from an invalid negation fixture syntax, corrected in later runs; the original artifacts remain preserved.
• Visibility-controlled repetitions retained identical visibility coverage throughout.
• Final verification found zero probe schemas and zero experiment sessions on either remote database.

The database experiments are complete. Full CI must pass before merge. These results do not establish TIN/Stannum performance parity or reveal TIN's compression codec or SIMD implementation.