fix(hindsight): complete recall tool methods

[ThomasMarcelis]

Summary

• Completes hindsight_recall support for method="recall", method="list", and method="entity".
• Adds validated per-call controls for budget, max_tokens, types, tags, tags_match, tag_groups, metadata filtering, list pagination, and entity token budget.
• Uses the current provider operation wrapper and the public Hindsight client.memory.list_memories(...) API.
• Fixes entity formatting for the current Hindsight client model shape and narrows scoped list/entity output so unrelated or untagged results do not leak through filtered calls.
• Falls back to default recall for unknown method values, matching existing tolerant control handling.
• Updates the Hindsight plugin README and user-facing memory-provider docs.

Investigation

• Current main uses _run_hindsight_operation(...) for Hindsight calls, so this keeps recall/list/entity behavior inside that provider operation wrapper.
• Raw/list retrieval uses the public client.memory.list_memories(...) path instead of private client internals.
• Current Hindsight entity responses expose RecallResponse.entities as a mapping of entity ids to entity state objects with entity_id, canonical_name, and observations; entity output now formats those fields directly.
• Metadata, tag, and type filters are applied after retrieval where the public API cannot express the full filter, including exact metadata scoping and exclusion of untagged list items when tags are requested.
• List item type detection accepts type, fact_type, and factType so current and older response shapes filter consistently.
• Invalid per-call controls fall back to configured defaults or safe bounds instead of breaking recall.

Validation

• python3 -m py_compile plugins/memory/hindsight/__init__.py tests/plugins/memory/test_hindsight_provider.py -> passed
• HERMES_HOME=/tmp/hermes-pr18268-provider PYTHONDONTWRITEBYTECODE=1 python3 -m pytest tests/plugins/memory/test_hindsight_provider.py -q -o addopts= -> 110 passed
• HERMES_HOME=/tmp/hermes-pr18268-memory PYTHONDONTWRITEBYTECODE=1 python3 -m pytest tests/plugins/memory -q -o addopts= -> 156 passed
• python3 -m ruff check plugins/memory/hindsight/__init__.py tests/plugins/memory/test_hindsight_provider.py -> passed
• git diff --check -> passed
• HERMES_HOME=/tmp/hermes-pr18268-full PYTHONDONTWRITEBYTECODE=1 python3 -m pytest tests/ -q -> blocked before collection because this local environment lacks pytest-xdist for the repo default -n auto.
• HERMES_HOME=/tmp/hermes-pr18268-full PYTHONDONTWRITEBYTECODE=1 python3 -m pytest tests/ -q -o addopts= -> 29 collection errors, 4 skipped; errors are missing local optional/dev dependencies (acp, fire).
• Baseline comparison: the same full-suite command on clean origin/main also reports 29 collection errors and 4 skipped for the same missing dependencies, so this PR does not introduce those local full-suite failures.

Notes / Caveats

• Live PR head after this update: 8040bd900714855407972f66fb5bc1c2662af7f6.
• The PR branch is mergeable but still diverged from current main (ahead_by=4, behind_by=34 at verification time).
• This PR does not modify workflow files. If maintainers require the branch to be updated onto current main, an owner/maintainer token with workflow scope may be needed because upstream main has workflow-file changes.

[teknium1]

Thanks for expanding the Hindsight recall surface. The premise remains valid on current main: hindsight_recall still exposes only query and directly calls arecall (plugins/memory/hindsight/__init__.py:314-326, :1729-1754).

Problems

• The PR omits Hindsight 0.6.1's supported opinion fact type from _VALID_RECALL_TYPES (plugins/memory/hindsight/__init__.py:57 in PR head), so _normalize_recall_types drops a valid request at :373. Main pins that client version in pyproject.toml:159.
• The list filter rejects every untagged item at proposed plugins/memory/hindsight/__init__.py:437, including tags_match="any" and "all". Hindsight 0.6.1 documents those modes as including untagged entries; only the strict modes exclude them. This makes list behavior differ from normal recall.
• The entity default_types=["world"] at proposed :1724 is shadowed by current main's non-empty observation-only default (plugins/memory/hindsight/__init__.py:1344-1351), so the PR's base-era world-default test needs reconciliation during salvage.

Suggested changes

• Preserve the client-supported opinion type and add a regression test.
• Make list post-filtering preserve the documented strict/non-strict tag semantics, with coverage for all four modes.
• Resolve entity type defaults against current main's recall-type policy before transplanting the tests.

Automated hermes-sweeper review.

[teknium1]

hindsight-client==0.6.1 also supports the opinion fact type for arecall. Dropping it here means a valid per-call types: ["opinion"] silently falls back; please include it or intentionally document and test a Hermes-level restriction.

[teknium1]

This makes any and all reject untagged memories, but Hindsight defines only any_strict and all_strict as excluding untagged results. Preserve those non-strict semantics so list filtering matches normal recall.

[teknium1]

On current main, _recall_types defaults to ["observation"], so _build_tool_recall_kwargs() selects that before this fallback. Reconcile this entity default and its test with the current observation-only policy during salvage.