Package EigenRegex for import (closes #13)

[InauguralPhysicist]

Closes #13.

Problem

EigenRegex was a flat five-file load_file library: its private helpers (_peek, pos, n, branches, …) lived in the global namespace and silently clobbered any consumer using the same names — exactly what eigen-sheet hit wiring regex find & replace (#36): two _peek definitions, last-one-wins, len: null has no length surfacing far from the cause.

Fix

Fold the split into a single importable root regex.eigs + eigs.json (name regex), matching the package-template contract (one root leaf; top-level non-underscore bindings become regex.*; _-names stay private). import regex now namespaces the whole engine, so the caller-globals collision class (#5) and the _peek clash (#13) are structurally gone rather than mitigated by local.

• Public surface: re_compile / re_match / re_search / re_find_all / re_replace (positional spans) + compat_match / compat_find / compat_replace (builtin-shaped, folded in from the compat shim — namespaced, so no builtin shadowing across the import boundary).
• Privatised the internal entry points to _rx_parse / _rx_compile_ast / _rx_vm_run; all other helpers were already _-prefixed.
• Tests converted load_file → import regex (S1–S8, smoke, bench). S9 flipped from the load_file scope suite to an import-isolation guard (now also survives a consumer _peek collision the old model couldn't). New tests/test_pkg_smoke.sh stages the package into eigs_modules/regex/ and imports it the way --pkg add would, asserting public surface + internal privacy.
• Docs (README / CLAUDE.md) moved to the import model; fixed EIGS_REF drift (CLAUDE.md said v0.26.0; actual pin is v0.30.0).

Verification

359 checks + package smoke, all green against the CI-pinned v0.30.0 (built from the tag). --lint regex.eigs: no issues.

Side finding (logged in GAPS.md, not fixed here)

Running the S8 differential against a v0.32.0 runtime surfaced a builtin regex_match shape change: non-participating capture groups went from truncated (regex_match of ["b","(a)|(b)"] → ["b"] on v0.30.0) to null-filled (["b", null, "b"] on v0.32.0). Not an EigenRegex bug, but it will break the S8 differential on an EIGS_REF bump and is a breaking builtin-shape change worth an upstream decision.

🤖 Generated with Claude Code

[InauguralPhysicist]

Closing as already-landed, not as rejected — and the reason is a mistake on my side worth recording.

During the ecosystem-wide v0.30.0 → v0.32.0 pin sweep, this repo's working tree was checked out on package-import-13. My sweep script cut its bump branch from ambient HEAD rather than from a freshly fetched origin/main, so this branch's entire contents rode along inside PR #15 and were squash-merged into main under the message "Bump EigenScript pin v0.30.0 -> v0.32.0". GitHub's squash body inherited the closes #13 trailer, which is why #13 auto-closed.

So the packaging work IS on main (41cc66e) and is green — tests/run.sh passes 359 checks plus the package smoke, and CI passed on #15. What went wrong is the record, not the code: a feature landed unreviewed under a pin-bump label.

This PR must not be merged now. Its branch predates the pin bump, so the only remaining delta against main is EIGS_REF/CLAUDE.md going v0.32.0 → v0.30.0 — merging it would silently revert the pin.

Fix applied going forward: sweep loops now branch from a fetched origin/HEAD explicitly and assert the staged diff contains only the intended files, and every merged sweep commit gets verified with git show --stat.