fix(runtime): raise on same-kind bundles claiming one class name

[sagi5060]

What

PluginRegistry._scan (agentdeck/runtime/registry.py) collected discovered classes into
found[attr.__name__] = attr, so two bundles of the same kind exporting the same class
name deduped silently, in sorted bundle order. Copying agents/greeter/ to
agents/greeter-v2/ to iterate and forgetting to rename the class made the original
greeter bundle vanish from the registry with no error — every request routed to that name
silently ran the other agent. Fixed at the point of discovery: _scan now tracks which
bundle claimed each class name and raises ConfigError naming both bundle paths and the
shared class name the moment a second, different class claims a name already taken.
AgentRegistry, WorkflowRegistry, App.load(), and InvocableRegistry (which discovers
through the same v1 scan) all inherit the fix for free.

Sibling nit from the same review (#82's Notes): a bundle whose module raises while
importing (SyntaxError, a missing dependency) used to surface as a raw traceback through
the import machinery, 12–14 frames deep. The import call in _scan is now wrapped, so it
raises ConfigError naming the offending bundle path with the original exception chained
as the cause.

Refs #82 — Done-when item 2 ("import/build failures wrapped") is only half done: import is
wrapped, build()/build_graph() is not (see judgment ledger and #119 below). Not closing
#82 until that's tracked to completion or explicitly descoped there.

Judgment ledger

• Raise, not warn. The issue asks to decide deliberately between a hard ConfigError and
  a loud warning that keeps both under distinguished names. A silently missing agent (today's
  bug) and a silently renamed agent (auto-disambiguating a warn-and-rename scheme) are the
  same failure mode from a caller's perspective — a request that used to reach one agent now
  silently reaches a different one, or an invented name nobody asked for. Raising is the only
  option where a name-routing mistake can't hide. I looked for a legitimate layout the loud
  version breaks and did not find one: v1 already keeps agents and workflows in disjoint
  registries and disjoint type_dirs, so two unrelated bundles that happen to pick the same
  class name for different concepts is exactly the case this issue exists to catch, not a
  case to accommodate.
• Collision check is identity-based, not name-based (fixed in review round 2 — see
  below): vars(module) yields one entry per binding, not per class, so a bundle that
  aliases its own class under a second name (GreeterAgent = Greeter, kept after a rename)
  must not trip the guard against itself. The check is claimant is not None and claimant is not attr, not attr.__name__ in found.
• This changes v1's frozen discovery error surface, as flagged in the issue. It is
  additive on the success path (a project with no name collisions, and no cross-bundle
  aliasing, behaves identically) and turns one specific silent-data-loss case into a
  load-time ConfigError. One existing test
  (tests/test_app.py::test_injected_session_factory_closed_when_load_fails) pinned the old
  raw RuntimeError a broken bundle's import used to raise; updated to expect the new
  ConfigError wrapping, since that behavior change is Notes item 2 of this same issue. No
  golden/SSE fixture exercises a same-kind collision, an alias, or an import failure, so
  nothing else moved.
• A ConfigError a bundle itself raises at module scope (not just SyntaxError/missing
  dep) is now re-wrapped too — the import except Exception is broad by design, so a
  bundle author's own ConfigError raised during import gets caught and re-raised as a new
  ConfigError naming the bundle path, with the original chained as __cause__. Type and
  message are preserved (nothing swallowed); the outer message is a superset (bundle path +
  str(original)), so no caller matching on the message's start breaks — checked, nothing in
  the codebase does.
• Scope of "import/build failures are wrapped": wrapped the importlib.import_module
  call inside _scan — the failure mode the issue's Notes describes concretely
  (SyntaxError, missing dep, a raw traceback through import machinery). Did not thread
  bundle-path context through BaseAgent.build() / BaseWorkflow.build_graph() — those are
  called from several sites (App.load(), InvocableRegistry.load(), run_agent, …) with no
  bundle-path association surviving past the scan (PluginRegistry.list() returns
  dict[str, type[T]], not a class-to-path mapping), and build() already has its own
  ConfigError paths for user-facing misconfiguration. Threading bundle paths through every
  build() call site is a larger, separately-scoped change — filed as discovery: build()/build_graph() failures still surface a bare exception with no bundle path #119 rather than
  done opportunistically here, and the PR now says Refs #82 instead of Closes #82 because
  of it.
• Message wording: f"two bundles under '{self.type_dir}/' both define the {self.label} class {attr.__name__!r}: '...' and '...'; one name is one invocable — rename one of the classes." — matches the lowercase, semicolon-joined style of the existing cross-kind
  message in agentdeck/runtime/discovery.py's _add rather than inventing a new voice.
• PluginRegistry's class docstring now states the collision/alias behavior — it
  previously documented only laziness and caching, so the new failure mode (and the
  non-failure alias case) wasn't discoverable from the class itself.
• CHANGELOG: kept the two new Unreleased "Fixed" entries. Reverted the edits to the
  two already-released 2.0.0b4 entries (from PR feat(registry): discover invocables into an InvocableRegistry #81) that stated "two bundles of the same
  kind ... still collapse to a single invocable" — verified via gh release view v2.0.0b4
  that the published release notes still say this, so a b4 reader was told the truth as of
  b4 and that section is the published artifact, not a living description of current
  behavior (the file's own header: entries are "written to be attached to a release as-is").
  dev's CHANGELOG now matches the release verbatim for that section; only the new
  Unreleased entries describe today's behavior.

Test evidence

Tests in tests/test_errors.py and tests/test_app.py:

• test_two_same_kind_bundles_sharing_a_class_name_raise_naming_both — two agent bundles
  (greeter, greeter-v2) both defining class Greeter; asserts the exact quoted bundle
  paths ('agents/greeter' and 'agents/greeter-v2', not bare substrings — one is a
  substring of the other, which is exactly how the previous version of this test passed
  even when the message named only one bundle) both appear in the ConfigError message.
• test_one_bundle_aliasing_its_own_class_is_not_a_collision (new, review round 2) — one
  bundle with GreeterAgent = Greeter; asserts App().load() succeeds with exactly one
  invocable. Pins the false-positive fix.
• test_bundle_import_failure_is_wrapped_with_its_path — a bundle that raises RuntimeError
  at import now raises ConfigError naming its path, with the original exception chained.
• test_injected_session_factory_closed_when_load_fails (existing, updated) — same import
  repro, now asserting the wrapped ConfigError instead of the raw RuntimeError.

Verified all of these have teeth by stashing the registry.py fix, rerunning — the
collision, alias, and import-wrapping tests all failed against the pre-fix code — then
restoring the fix and rerunning green.

make check equivalent (fresh python3.12 -m venv + pip install -e ".[dev,serve,durability]",
since the shared repo .venv's editable install resolves to the main worktree, not this
one): ruff check — pass; ruff format --check — pass; ty check agentdeck — pass;
lint-imports — 12/12 contracts kept; pytest tests/ — 688 passed, 96 skipped (expected:
optional Redis/Postgres backends without services running), 0 failed. Re-ran after merging
origin/dev in (brought in #112's control-lifecycle events, unrelated to this PR;
CHANGELOG.md conflict resolved as a union) — still fully green.

Review round 2

Addressed request-changes review:

1. (blocking, fixed) Collision check was name-based, so a single bundle aliasing its own
   class (GreeterAgent = Greeter) tripped the guard against itself with a
   self-contradicting message naming the same bundle path twice. Now identity-based
   (claimant is not attr). Added the false-positive regression test.
2. (blocking, fixed) The collision test's assertions were bare substrings
   ("agents/greeter" in message), satisfied even if the message named only
   agents/greeter-v2. Now asserts the exact quoted forms, which are not substrings of each
   other.
3. (blocking, fixed) Restored the two already-released 2.0.0b4 CHANGELOG entries
   verbatim — dev and the published v2.0.0b4 release notes now agree again. Only the new
   Unreleased entries describe current behavior.
4. (advisory, acted on) Filed discovery: build()/build_graph() failures still surface a bare exception with no bundle path #119 for the build()/build_graph() half of "import/build
   failures are wrapped", and changed Closes #82 to Refs #82 since that Done-when item is
   only half satisfied by this PR.

Also took the wording/docstring advisories and added the re-wrapped-ConfigError ledger
line above.

🤖 Generated with Claude Code

[sagi5060]

Review round 2 addressed — pushed to this branch, ready for re-review.

1. (blocking, fixed) Collision check was name-based (attr.__name__ in found), so a
   bundle aliasing its own class under a second name (GreeterAgent = Greeter) tripped the
   guard against itself with the self-contradicting message you found
   ('agents/greeter' and 'agents/greeter'). Reproduced it, then switched to identity
   (claimant is not attr) and added test_one_bundle_aliasing_its_own_class_is_not_a_collision,
   which fails on the pre-fix code with exactly that message.
2. (blocking, fixed) test_two_same_kind_bundles_sharing_a_class_name_raise_naming_both
   now asserts the exact quoted forms ("'agents/greeter'", "'agents/greeter-v2'") instead
   of bare substrings. Verified by mutating the message to drop bundle_of[...] (naming only
   the second bundle) — the old assertions still passed, the new ones fail.
3. (blocking, fixed) Restored both 2.0.0b4 CHANGELOG entries verbatim, matching
   gh release view v2.0.0b4. Only the new Unreleased entries describe current behavior; dev
   no longer disagrees with the published release notes.
4. (advisory, acted on) Filed discovery: build()/build_graph() failures still surface a bare exception with no bundle path #119 for the build()/build_graph() half of "import/build
   wrapped with the bundle path" (deliberately out of scope here — no bundle-path association
   survives past PluginRegistry.list(), and it touches several call sites, not one choke
   point). Changed Closes #82 to Refs #82 accordingly.

Also took the wording (agentdeck/runtime/discovery.py-style message), docstring, and
ConfigError-re-wrap ledger advisories.

make check equivalent green after merging origin/dev (brought in #112, unrelated):
ruff, ruff-format, ty, lint-imports (12/12 kept), pytest tests/ — 688 passed, 96 skipped, 0
failed.