perf(ir): retain *runtime-defn-ir-cache* for only the current ns

[nnunley]

Summary

IR pipeline cache retention: retain runtime-defn-ir-cache for only the current namespace.

When enable-inline is on, the cache stashes each defn's built IR for same-ns inline registry seeding. The registry only seeds same-ns, so entries for other namespaces are dead weight. This change makes each seed replace the cache with a single-key map {current-ns {...}} instead of accumulating across all loaded namespaces.

Decomposition Note

This PR has been decomposed per reviewer feedback:

• compile-def-fn-value (def+name* IR seam) → moved to feat(ir): def+name* IR-compile seam — top-level (def x (fn …)) via the IR path #647
• Test files (ir_blockarg_cond.lg, ir_junk_agreement.lg, ir_and_cond_rpo.lg) → moved to fix(ir/lower): block-junk agreement + RPO block order + deferrable-branch guards #648

This PR now contains the cache-window retention change only — the isolated performance improvement for same-ns inline caching.

Base & Merge Order

Base: main@upstream
Merge order: 4th (lands after #648, #649, #647 in stack)

Verification

• ✓ Build: go build ./... passes
• ✓ Tests: go test ./pkg/ir passes
• ✓ Generated: make check-generated passes

[mparrett]

The body says self-contained to pipeline.lg, but this PR carries two things belonging to its siblings:

• compile-def-fn-value — the Lisp half of #647's Go hook. #647 resolves ir.passes.pipeline/compile-def-fn-value and no-ops when the lookup fails; that var isn't on main, and it's defined here with no caller inside this PR. So #647 merged alone is inert, and this one merged alone defines a function nothing calls. I verified both directions by building them.
• test/ir_junk_agreement.lg and test/ir_and_cond_rpo.lg — headed "#625 fix 1" and "#625 fix 2", the block-junk-agreement and RPO fixes that live in #648. I ran both against origin/main: 7 and 9 passing. They're parity tests rather than repros, so they don't refute #648's "could not reproduce" note — but they're named for #648's fixes and belong with them, and their headers need adjusting either way (details on #648).

So the four-way split doesn't decompose the way the descriptions say. All four target main and all four report MERGEABLE, which hides it. Suggested regrouping:

• #647 plus the compile-def-fn-value defn — one landable seam change, with whatever coverage turns out to be possible
• #648 plus the two tests above and the baseline rebaseline
• #649 retargeted onto #648's branch
• #650 reduced to the retention change itself

The retention change reads well on its own. The single-ns window is backed by the code rather than only the comment — seed-inline-registry-from-cache reads (get @*runtime-defn-ir-cache* ns-name {}) at pipeline.lg:66 and its only caller passes the current ns at :694 — so entries for other namespaces really are dead weight. The "a require mid-file switches the cache to the dependency's ns and back" caveat is the case I'd have asked about, and naming the cost as a forfeited inline opportunity is the right read. No objection to that part once it's on its own.

[mparrett]

Re-reviewed the isolated cache-window change at the current head. The implementation matches the same-namespace lookup contract, the tests pin same-namespace accumulation and cross-namespace eviction, targeted local tests pass, and all required checks are green. Approved.