Add the allocation gate: alloc-diff and assert --max-alloc-mb

[jantimon]

Stacked on #213 (assert --format json). Review only this PR's own diff; merge from the tip of the stack. The alloc gate is wired through #213's AssertView, hence the stack.

Problem

A GC-pressure regression that allocates hard but costs little CPU passed every wpd gate. In the dogfood, cpu-diff caught a +61% allocation regression only by accident (building the array also cost CPU); a pure large-buffer / retained-churn regression sails through counts, timing, slice, and CPU gates alike.

Change

Two owed surfaces, mirroring the existing CPU gate:

• alloc-diff <baseline> <current> --fail-on-regression gates on net allocated bytes (the allocation analog of cpu-diff's net JS self-time), with per-package and per-function byte movers, a comparability refusal across an incompatible workload/lane/capture, and an AllocDiffResult under --format json|toon.
• assert --max-alloc-mb <mb> gates the total sampled MB against a budget.

Both read a --target node --alloc recording's sidecar .alloc.json. A recording with no alloc model (a chrome or node-cpu capture) is a loud n/a-FAIL, never a silent pass; a run-group (no --alloc member exists) is an n/a-FAIL too.

The noise floor (measured, and consistent with the cpu-diff fix)

Sampled byte totals are the directional tier (~10-20%), so the gate floor is percentage-led, the same two-term shape as cpu-diff: net must clear max(--noise-floor MB, --noise-pct% of the baseline), default max(1 MB, 25%).

[measured, --target node --alloc, gate lock held] on a ~14 MB workload (examples/probes/allocates.mjs at --iterations 20, same module path both sides):

	identical-code net	+50% regression net
relative	p50 6%, p95 15%, max 20%	+33..78%

25% sits above the 20% identical-code ceiling with margin. Verified through the built CLI's exit code:

• identical pairs: 0/90 false-red (100% green)
• +50% regression pairs: 80/80 caught (100%)
• 1 MB absolute floor covers a tiny-allocation workload where 25% would be a fraction of a sample; --noise-floor / --noise-pct tune each term.

Scope

• src/commands/allocdiff.ts (new), src/model/query.ts (AllocDiffResult + delta types), src/cli.ts (alloc-diff + --max-alloc-mb), src/commands/assert.ts (the alloc axis, n/a-FAIL without a model), src/index.ts + README type table + public-types walker.
• Docs: allocation-profiling.md gains a gating section with the measured floor; facts.md pins the 25% default; README CI-gating section gains the alloc gate + a baseline-management paragraph (where baselines live, what is lean to commit, the re-record-on-upgrade rule).
• Tests: test/unit/alloc-diff.test.mjs (6: identical, scaling floor, +50% gate, absolute-floor, knob, comparability refusal) and 2 assert alloc-axis tests (gates the sibling total; n/a-FAIL without a model).

Gates: lint, format:check, build, unit (753), measurement (6), the 3 Chrome assert e2e tests all green.

Replaces #214, which was auto-closed during a botched restack after its base (#213) squash-merged; this is the same change rebased onto main with the README gating paragraph merged against #212's noise-floor text. Unit suite green on the restacked head (759 pass).