fix: hold meta-page lock across reinit and writes in MetaPage::store

[pavanmanishd]

Summary

Parallel DiskANN index builds can fail with assertion 'left == right' failed (see #264). Every parallel worker calls MetaPage::store(false) from finalize_index_build (build.rs:940), and store() released the exclusive buffer lock between reinit and each subsequent write. Concurrent workers could interleave the operations and produce the observed offset mismatch.

This PR restructures MetaPage::store to hold one WritablePage across reinit and both writes, committing once at the end. Concurrent callers now serialize on the Postgres buffer lock.

Note for reviewers: the root cause is slightly different from what's described in #264 — the MetaV1 branch of fetch() is not the trigger. Freshly built indexes use PageType::Meta, and fetch() on that branch is read-only (Self::load). The race I'm seeing is in MetaPage::store itself, exercised from finalize_index_build in each parallel worker.

Changes

• pgvectorscale/src/access_method/meta_page.rs — store() now acquires a single WritablePage (via new for first_time=true, or modify + in-place reinit for first_time=false), writes header + meta, and commits once.
• pgvectorscale/src/util/chain.rs — new write_chain_item_to_page(page, data) helper for the single-item, caller-holds-lock case (keeps the ChainItemHeader serialization format in one module). Removed the now-unused ChainTapeWriter::reinit.

Net diff: +33 / −34 lines across 2 files.

Test plan

• cargo pgrx install --release --features pg18 builds cleanly; no new clippy warnings on the changed files.
• Reproduced the race on main with a 50 ms sleep inserted between reinit and the first write in store(): 4 of 5 iterations fail with the exact assertion from [Bug]: <Title>Parallel_Build_Assertion_Failure #264.
• Verified the fix — with the patch applied and a 50 ms sleep inserted inside the single lock window to stress serialization: 5 of 5 iterations pass.
• cargo pgrx test pg{17,18} — to be confirmed by CI (local run blocked by a pgrx --sudo install quirk on macOS, unrelated to this change).

Repro script preserved for convenience:

DROP TABLE IF EXISTS v;
CREATE TABLE v (id SERIAL PRIMARY KEY, embedding vector(64));
INSERT INTO v (embedding)
SELECT (SELECT array_agg(random()::real) FROM generate_series(1, 64))::vector
FROM generate_series(1, 50000);

SET diskann.min_vectors_for_parallel_build = 100;
SET max_parallel_maintenance_workers = 16;
SET diskann.force_parallel_workers = 16;

DO $$ BEGIN FOR i IN 1..5 LOOP
    BEGIN
        EXECUTE 'DROP INDEX IF EXISTS v_diskann_idx';
        EXECUTE 'CREATE INDEX v_diskann_idx ON v USING diskann (embedding vector_l2_ops)
                 WITH (storage_layout = ''memory_optimized'')';
        RAISE NOTICE 'Iteration % succeeded', i;
    EXCEPTION WHEN others THEN
        RAISE NOTICE 'Iteration % FAILED: %', i, SQLERRM;
    END;
END LOOP; END $$;

Notes

I intentionally did not add an in-tree regression test because reproducing the bug deterministically requires injecting a sleep inside store(). Happy to add a #[cfg(test)]-gated race-widener test on request.

Fixes #264

[CLAassistant]

All committers have signed the CLA.