chore(deps): bump pydantic-settings, transformers, soupsieve (security)

[dcbouius]

Closes the pydantic-settings Dependabot alert across every affected manifest, plus the three high-severity alerts in the root lock.

Package	From	To	Advisory
pydantic-settings	2.12.0 / 2.14.0 / 2.14.1	2.14.2	GHSA-4xgf-cpjx-pc3j
transformers	5.3.0	5.12.1	GHSA-fgcw-684q-jj6r
soupsieve	2.8	2.8.4	GHSA-2wc2-fm75-p42x, GHSA-836r-79rf-4m37

pydantic-settings is transitive everywhere (no direct declaration), so the locks are the only lever. It's bumped in all 7 flagged manifests: root, hindsight-integration-tests, strands, openai-agents, dify, claude-agent-sdk, agno.

Reviewer notes

1. The transformers floor bump is a behavioral change for downstream consumers.

transformers is a direct dep and hindsight-api is published to PyPI, so the declared floor — not our lock — is what protects installers of the local-ml / local-onnx extras. The old >=4.53.0 floor resolves to 4.57.6 (vulnerable) under any downstream cap of transformers<5, so a user could get a vulnerable install while our lock looked clean. Raised to the advisory's first patched version, >=5.5.0.

The trade: consumers pinned below transformers 5 now get a resolution error instead of a silent vulnerable install. That's intended, but it's the notable call in this PR.

Worth knowing the old floor was already unreachable in practice — 4.53.0 requires tokenizers<0.22, which our own cap excludes.

2. The tokenizers<=0.23.0 cap is deliberately kept.

It looks redundant (every transformers >=5.5.0 declares the same cap itself), but it is the guard for #2055: transformers shipped metadata declaring a wider tokenizers range than its import-time check enforces, which let 0.23.1 in and broke local embeddings at startup. Dropping the cap re-opens that door and trusts the metadata that already failed once. The comment now records this so it doesn't read as removable next time.

3. crewai is intentionally not bumped.

Its lock has pydantic-settings 2.10.1, which looks stale next to this sweep. The advisory's range is >=2.12.0, <2.14.2 — NestedSecretsSettingsSource didn't exist in 2.10.x, so it's genuinely unaffected. Dependabot correctly did not flag it.

4. Root uv.lock diff is 3353 lines; only 3 are real.

uv rewrites in its own current format whenever it writes, so the upgrade dragged the root lock from revision 1 to 3 and added upload-time metadata throughout. This isn't drift — the other 32 locks in the repo are already revision 3 and CI's setup-uv@v7 is unpinned, so this brings root in line. It couldn't be split into a separate reformat commit because a no-op uv lock doesn't write at all. Verified the change set contains exactly 9 version changes and nothing else.

Verification

#2055 failed at import, not resolution, so this was checked past the lock diff:

• uv sync --extra local-ml resolves tokenizers 0.22.2 under transformers 5.12.1
• LocalSTEmbeddings initializes and returns 384-dim vectors matching its declared dimension
• LocalSTCrossEncoder initializes and scores sanely (relevant 9.55 vs irrelevant -11.17)
• ./scripts/hooks/lint.sh passes
• Local-ML test files: 30 passed, 10 skipped (9 errors are connection refused on port 5432 — no local Postgres, unrelated)

🤖 Generated with Claude Code

[dcbouius]

Rebased onto current main (154998701, includes the v0.8.4 release). The CI failures were all from the stale base, not this change:

• test-api — test_db_abstraction.py Oracle session-schema tests (_FakeAsyncCursor has no attribute 'fetchone', cursor() must not be called for the public schema): fixed by fix(test): green main — update Oracle session-schema tests for #2708 reset behavior #2722 (fix(test): update Oracle session-schema tests for #2708 reset behavior), which landed after this PR's original base. The tests were renamed/updated there; both pass locally on the rebased tree.
• test-embed-windows — issue daemon_embed_manager: prefer locally-installed hindsight-api over uvx #1240 uvx-fallback regression: fixed by fix(embed): green main — use --target sibling binary unconditionally (#2676 vs #1240) #2723 (fix(embed): use --target sibling binary unconditionally), also post-base.
• Core LLM tests — test_date_field_calculation_yesterday: an LLM-as-judge temporal test, non-deterministic by design (see CLAUDE.md testing conventions). Unrelated to a dependency bump; should clear on re-run.

Re-verified after rebase: the root uv.lock re-locks as a no-op with exactly the 3 intended version changes (pydantic-settings 2.14.2, transformers 5.12.1, soupsieve 2.8.4), workspace versions correctly reflect v0.8.4, and the local-ML import/embedding/reranker path still resolves tokenizers 0.22.2 under transformers 5.12.1.

[dcbouius]

The remaining CI reds are environmental, not caused by this dependency change:

• test-api (2/3) — 46× ImportError: sentence-transformers is required + Failed to start embedded PostgreSQL. sentence-transformers is 5.2.0 in this branch's uv.lock, byte-identical to main — this bump does not touch it, and transformers 5.12.1 + sentence-transformers 5.2.0 was verified locally (LocalSTEmbeddings returns 384-dim vectors; LocalSTCrossEncoder scores correctly). The job's venv cache key hashes uv.lock + pyproject.toml, so this PR is a cache miss; shards restore a stale venv via restore-keys and reconcile with uv sync --frozen --all-extras. Shards 1 and 3 pass; only shard 2 ends up without sentence-transformers, alongside an embedded-Postgres start failure — a sharded-venv-cache/infra flake, not a dependency defect.
• Core LLM tests — LLM-as-judge non-determinism (test_date_field_calculation_yesterday, test_cognitive_epistemic_dimension, test_reflect_chinese_content) plus a Gemini 400 INVALID_ARGUMENT ("Multiple tools are supported only when they are all search tools") provider error. Unrelated to a lockfile bump.

The dependency change itself is verified: root uv.lock re-locks to a no-op with exactly the 3 intended version changes, npm ci/local-ML import path all clean. These jobs may need a maintainer re-run on a warm cache, or a look at the sharded-venv caching if shard 2 keeps missing --all-extras deps on cache-miss PRs.