Store page pointers as dense ordinals (format 1.3)

[hadashiA]

Every on-disk page pointer (index roots, node siblings, internal-node children, overflow blob refs, secondary-index PageRefs) now stores a dense page ordinal assigned in flush order instead of a byte offset. A page directory section at the end of the file maps ordinal -> offset (~0.2% of file size at 4 KB pages), back-patched into the header.

The reader's page cache is indexed directly by ordinal:

• the ConcurrentDictionary map becomes a plain Entry?[] — a cache hit is one volatile array read + a refcount increment; no hashing, no key comparison, no per-entry node objects
• publish/evict become single CAS operations on the slot
• the ghost set becomes an eviction-epoch int[] instead of a dictionary (allocation-free, exact FIFO window)

IPageLoader now takes the byte offset directly (translated through the directory only on cache misses), so loader implementations are unchanged in behaviour.

Breaking format change: the builder always writes 1.3 and the reader accepts 1.3 only — older files fail with an explicit StorageFormatException asking for a rebuild. Key digests (1.1) and Eytzinger layout (1.2) remain per-page flags, orthogonal to this.

Benchmarks (same session, Apple M5 / .NET 10, per 1000 lookups)

Method	before	after	delta
FindByKey (10k)	19.99 us	14.89 us	-25.5%
FindByKey_RandomKeys	27.27 us	22.71 us	-16.7%
FindByKey_RandomKeys_NoRepeat	58.87 us	50.92 us	-13.5%
FindByKey_ParallelSpread (8T)	132.12 us	108.14 us	-18.2%
FindByKey_1M	31.59 us	23.96 us	-24.2%
FindByKey_1M_RandomKeys	80.57 us	56.01 us	-30.5%

Open + first read also drops from 117.6 us / 652 KB allocated to 33.2 us / 19 KB, since the cache no longer preallocates dictionary buckets.

All 85 tests pass. The README benchmark table/charts are intentionally left as-is in this PR (the numbers there come from a different machine); they can be refreshed in one sweep later.

🤖 Generated with Claude Code