Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@CHANGELOG.md`:
- Line 18: Update the CHANGELOG bullet that currently reads "No token limits or
temperature controls — uses provider defaults" to accurately reflect that
Anthropic's max_tokens is hardcoded in the codebase; specifically replace or
append that line to mention "Anthropic: max_tokens is hardcoded" (or similar
wording) so users are not misled that all providers use defaults, referencing
the existing bullet text and the Anthropic max_tokens configuration.

In `@README.md`:
- Line 738: Remove the roadmap-status sentence starting with "The features on
the roadmap — \<Http\> ... All four pieces are now complete." from README.md and
instead add or update that status entry in ROADMAP.md (preserving the content
and any strike-throughs/issue links); then replace the removed sentence in
README.md with a short link/reference to ROADMAP.md (e.g., "See ROADMAP.md for
current feature status") so README only links to the canonical roadmap and
ROADMAP.md holds the authoritative status text.
- Line 66: The README line claiming "mockable" for LLM inference is too broad;
update the phrase referencing Inference.complete to qualify that mocking is only
supported on the host/test side (e.g., "host-side test mocking") and explicitly
note that handle[Inference] is still unsupported until issue `#372`, or remove the
"mockable" claim entirely until `#372` lands; specifically edit the sentence
mentioning Inference.complete to either append a parenthetical like "(host-side
test mocking only; handle[Inference] unsupported until `#372`)" or replace
"mockable" with a qualified phrase as described.

In `@ROADMAP.md`:
- Line 23: Update the v0.0.100 summary to remove or correct features that
actually belong to v0.0.101: specifically, detach the "<Inference>" effect and
the "64-program conformance suite" from the v0.0.100 sentence and either move
them to a v0.0.101 entry or rephrase the v0.0.100 line to exclude those items;
edit the text referencing "v0.0.100" so it only lists features that truly land
in that release and add a separate v0.0.101 bullet containing "<Inference>" and
the 64-program conformance count if needed.

In `@SKILL.md`:
- Around line 1089-1099: safe_classify calls classify(`@String.0`) without
ensuring classify's precondition string_length(`@String.0`) > 0; fix by either
strengthening safe_classify's requires clause to require
string_length(`@String.0`) > 0 or by guarding the call (check
string_length(`@String.0`) > 0 and only call classify when true, returning
"unknown" otherwise); reference the functions safe_classify and classify and the
predicate string_length(`@String.0`) > 0 when making the change.

In `@spec/09-standard-library.md`:
- Around line 525-565: Examples in the spec still treat Inference.complete as
returning a plain String but its signature is Result<String, String>; update
every example (notably the classify snippet and any lines that use "let `@String`
= Inference.complete(...)" or similar) to handle the Result type by binding the
result (e.g., let `@Result` = Inference.complete(...)) and then pattern-matching
or unwrapping (Ok(completion) / Err(message)) before using the completion;
ensure functions like classify either propagate the Result (return
Result<String,String>) or explicitly match and return the inner String on Ok or
propagate/convert Err, and apply the same change to all downstream examples that
call Inference.complete so they are valid Vera code against the new API
contract.

In `@tests/test_codegen.py`:
- Around line 9615-9623: The test currently calls os.environ.clear() then
os.environ.update(orig) in the finally block (around the execute(result) call
and the orig snapshot), which wipes the entire process environment and can break
parallel tests; instead, restore only the keys you changed: compute the set of
keys added during the test (set(os.environ) - set(orig)) and delete those, and
for keys present in orig restore their original values via
os.environ.update(orig) or use a test helper/monkeypatch to set only VER
A_ANTHROPIC_API_KEY; replace the os.environ.clear()/update(orig) sequence with
targeted removal/restoration of modified keys (reference the orig variable, the
os.environ.clear()/update(orig) calls, and the execute(result) test block to
locate where to apply the change).
- Around line 9566-9660: Add unit tests in TestInferenceCollection to exercise
provider/model selection and overrides: create tests that (1) simulate OpenAI
and Moonshot auto-detection by patching environment and the internal
_call_inference_provider to verify the chosen provider path (e.g.,
test_inference_auto_detect_openai, test_inference_auto_detect_moonshot), (2)
assert that setting VERA_INFERENCE_PROVIDER forces the provider used, and (3)
assert that VERA_INFERENCE_MODEL overrides the model selection; use the existing
patterns in test_inference_complete_mocked_success/failure for patching
"vera.codegen.api._call_inference_provider", manipulating os.environ, compiling
_CLASSIFY_SOURCE via _compile_ok and executing via execute, and restore
os.environ in finally blocks so tests clean up after themselves.

In `@vera/codegen/api.py`:
- Around line 170-172: The HTTP calls use _urlreq.urlopen without a timeout and
should be bounded: add a finite default timeout constant (e.g.,
DEFAULT_PROVIDER_TIMEOUT) and pass it into _urlreq.urlopen(req,
timeout=DEFAULT_PROVIDER_TIMEOUT) for the POST call that builds req via
_urlreq.Request; wrap the call in the same error handling used elsewhere so
timeout/socket/URLError exceptions are caught and returned as Err with an
expiry/timeout marker (use the existing Err type/symbol), and apply the
identical change to the other two provider branches that also call
_urlreq.urlopen (the occurrences around the other Request/urlopen blocks).
- Around line 2016-2041: host_inference_complete currently reads _os.environ
directly (VERA_INFERENCE_PROVIDER, VERA_ANTHROPIC_API_KEY, VERA_OPENAI_API_KEY,
VERA_MOONSHOT_API_KEY, VERA_INFERENCE_MODEL) which ignores the
execute(env_vars=...) sandbox; change it to resolve env via the same mechanism
as host_get_env or the env dict passed through execute, e.g., replace
_os.environ.get(...) calls with the host_get_env(key, default) helper or lookup
in the execute-provided env map and then use those values when selecting
provider and calling _call_inference_provider so tests and guest IO see the same
environment.

In `@vera/environment.py`:
- Around line 418-429: The Inference effect registration only defines the
"complete" operation but is missing the "embed" operation, so update the
effects["Inference"] EffectInfo to include an OpInfo entry for "embed" alongside
"complete"; add an "embed" key to the operations mapping (similar to how
"complete" is created with OpInfo) and supply the correct parameter and return
types per Issue `#61` (match the expected type signature used by the standard
library), ensuring the operation name "embed" and the EffectInfo/OpInfo usage
remain consistent with the existing patterns in effects["Inference"],
EffectInfo, and OpInfo.

---

Outside diff comments:
In `@vera/codegen/compilability.py`:
- Around line 43-52: The rationale message passed to self._warning for
error_code "E603" is stale: it lists compilable effects but omits "Inference"
which is now supported; update the rationale string in the same call (the branch
handling effects where eff.name != "Inference" and not handled) to include
"Inference" alongside "pure, IO, Http, State<T>, Exn<E>, and Async" so the
message accurately reflects current compilable effects (refer to the
self._warning call that uses decl, the f"Function '{decl.name}' uses unsupported
effect '{eff.name}' — skipped." message, and error_code="E603").

In `@vera/codegen/functions.py`:
- Around line 167-187: Two qcall handlers (the Http and Inference qualifier
cases) incorrectly set the private flag self._needs_alloc instead of the
canonical allocator flag self.needs_alloc; update the assignments in the Http
and Inference qcall handling code to set self.needs_alloc = True (replacing
self._needs_alloc = True) so they match other paths (arrays, strings, regex,
decimals, map, set, json, html) and align with the WasmContext propagation used
by functions.py and context.py.

In `@vera/wasm/calls.py`:
- Around line 365-374: In the host-call handling branches for call.qualifier ==
"Http" and "Inference" (the code that builds wasm_name like f"http_{call.name}"
and f"inference_{call.name}" and appends instructions with call
$vera.{wasm_name}), replace assignments to self._needs_alloc = True with
self.needs_alloc = True so the canonical allocator flag is set; make this change
in both the Http and Inference branches and keep the rest of the logic (adding
to self._http_ops_used / self._inference_ops_used and appending the call
instruction) unchanged.