Make WorkTable's own source std-free

[pathscale]

cargo check --no-default-features now passes with no errors and no warnings. The default build is unchanged and the same 926 tests pass.

What the numbers were

Counting std:: paths in the source predicts nothing useful, so the work was driven by the compiler instead. Flipping #![no_std] on produced 550 errors. They fell into three groups of very different character:

	count	what it was
core and alloc already have it	~170 paths	fmt, mem, hash, marker, cmp, ops, iter, num, ptr, hint, any, array, future, pin, task, cell, error, atomics, Arc, Rc, Duration, VecDeque, the BTrees
the alloc prelude	528 errors	Vec 356, String 68, vec! 47, format! 41, Box 26
genuinely wants an OS	22 errors	every one in the persistence half

Where the line actually falls

Not where the module tree suggested. persistence looked like the std half wholesale, but the table's own mutation path builds Operation values inline, so gating the module took the insert path with it. Those types are plain data: operation and operation::util name no std path, no filesystem and no runtime, and they stay in the core.

The gate falls on what wants an operating system: engine, error (which is a path and a reason), space, task, readonly_engine, operation::batch, and the PersistenceEngine trait that speaks in batches. vacuum goes with them, needing Instant for its settle deadline and PersistenceResult for its trait; a table with no background runtime has nothing to run a defragmenter on.

Also gated: validate_persisted_state (audits a loaded file, so takes a Path), the WorkTableError::Persistence variant, the MemStat impls for SystemTime and Instant, and migration.

Two changes that are not mechanical

Hash maps move to hashbrown, which is what std's own HashMap is built on. Its get is bounded on Equivalent<K> where std's is bounded on Borrow<Q>, so a key written &1.into() no longer infers and now names its type. BatchData is a public alias, so a caller building one with std::collections::HashMap must say worktable::prelude::HashMap; the prelude re-exports it.

Generated code cannot name alloc or hashbrown, since a user crate has neither in scope, so it goes through the prelude, and the crate gains an extern crate self as worktable alias so those paths also resolve where they are emitted.

What this does not do

It does not make the crate build for a bare-metal target. arc-swap has no std feature at all, our parking_lot fork deliberately drops the generic thread parker, and data_bucket is its own piece of work. Each was measured separately by building it as its own #![no_std] crate for aarch64-unknown-none.

What it establishes is that none of the remaining distance is WorkTable's own code.

🤖 Generated with Claude Code

[pathscale]

Superseded by #102, which contains every commit on this branch: git merge-base --is-ancestor origin/feat/no-std-core origin/feat/tunable-page-stride passes. Reviewing them separately means reviewing the same diff twice, and #102 is where the measurements for both live. Closing so the repo has one PR for this work rather than a stack.