feat(parallel): let any number of same-class tasks be in flight at once

[lisachenko]

Closes #15 — Option 1, per-instance keys upstream. Depends on lisachenko/php-shared-data-extension#23; CI here stays red until that merges into the substrate's main (this repo requires it as dev-main).

What this does

$a = $runtime->spawnParallel(new RenderJob(1));
$b = $runtime->spawnParallel(new RenderJob(2));   // now just works

The substrate now registers a persisted graph under a name minted from its own root address (persistInstance()), so a second instance of one class is a second entry — never an upsert, never a release under a running worker's feet. With the hazard gone at the source, the runtime deletes its guard rather than keeping a softened version:

• SharedArena::persist() rides persistInstance(). Each task graph has its own registry entry; any number of one class are live at once.
• Shared roots are keyed by the declared name, not the class — fixing the adjacent break the class keying caused: two declareShared() roots of one class used to upsert each other's entries. New test pins two SharedCounter roots mutated independently.
• ArenaTaskDirectory drops the in-flight ledger, the refusal, and releaseInFlight(); WorkerSupervisor drops the dispatched map that existed only to feed it. The honest cost statement replaces the limit in the docblock and in AGENTS/README: an unpublished spawn clones its graph into the arena and that memory lives until family teardown — steady-state workloads publish before the fork, which allocates nothing per spawn.
• Each panic keeps its own SharedError. Previously a second worker's capture superseded the first worker's registry entry, so a waiter attaching the first error's address after that found nothing. New test drives two concurrently panicking workers and asserts each waiter reads exactly its own message.

Acceptance criteria from #15

• ☑ Two tasks of the same class in flight simultaneously, each await()s its own result — testTwoTasksOfOneClassAreInFlightAtOnce.phpt pins two napping same-class tasks to two workers so their lifetimes genuinely overlap.
• ☑ Neither task's arena graph is mutated or freed while the other is running — asserted, not assumed — both instances are persisted pre-fork at verified-distinct addresses, and while both tasks run, each graph's string payload is read back through shared memory and still carries its own value.
• ☑ The pre-fork publishTask() route keeps working unchanged — the entire existing suite drives it; no published-route code changed.

Verification

• Full suite OK (97 tests, 97 assertions) on PHP 8.4 against a vendor tree carrying the substrate branch (94 existing + 3 new).
• PHPStan level max clean; cs-check clean.
• The arena-watermark soak's RSS verdict fails identically on unmodified main (24 KiB at 40 rounds, scaling with rounds, arena metrics byte-flat) — a pre-existing parent-side native-memory accumulation, filed as soak-arena-watermark reports a steady parent-side RSS climb the arena metrics do not see #24 rather than folded in here.

🤖 Generated with Claude Code

https://claude.ai/code/session_01U5QRWPfmkZME7hDjU6i5EA

---

Generated by Claude Code

[lisachenko]

CI status, verified from the job logs: all four failures (Tests + PHPStan, both minors) are the declared dependency and nothing else — the runners resolved lisachenko/php-shared-data-extension at upstream dev-main 5d170a0, which predates persistInstance(). PHPStan reports exactly the three errors that implies (persistInstance() not found, plus the old class-string signature of persist()).

Unblock path: merge lisachenko/php-shared-data-extension#23 (green, 8/8 jobs), then CI here re-runs against the new dev-main and should match the local result — 97/97 tests, PHPStan level max clean. I'll re-trigger the run as soon as the substrate merges.

---

Generated by Claude Code