fix(memory): enforce Hindsight routes across recall, reflect, and retain

[jackroofan]

Summary

• select one Hindsight route from conversation context and apply it consistently to automatic recall, explicit hindsight_recall, hindsight_reflect, and retain tagging
• enforce deterministic selection: exact chat_ids > exact case-insensitive chat_names > case-insensitive keywords in the query/chat name; same-signal ties use configuration order
• keep route filtering effective across clients with and without tag_groups, including exclusions, priority-first recall, deduplication, result caps, and one bounded type-broadening fallback
• make reflect honor recall_max_tokens, use routed filters, and fall back to routed recall when synthesis/tag-group support cannot provide an answer
• optionally suppress short low-signal recall while allowing configured domain keywords through

Configuration contract

recall_routes accepts either a JSON object or the JSON string emitted by the plugin config UI. A non-empty object enables routing. Each ordered named route may contain:

• selectors: chat_ids, chat_names, keywords
• recall filters: tags, tags_match, exclude_tags, query_prefix
• priority pass: priority_tags, priority_tags_match, priority_query_prefix
• fallback/result controls: min_results, max_results
• retain behavior: retain_tags

Global optional controls are:

• recall_max_results (0 means unlimited; a route can override it)
• recall_skip_low_signal_queries
• recall_low_signal_min_chars
• recall_domain_signal_keywords

Existing configs remain unchanged when these settings are absent. If no route matches, existing global recall_tags / recall_tags_match behavior is preserved.

Behavior and fallback

• Positive and negative route filters are sent as tag_groups when supported.
• If the client rejects tag_groups, recall retries with positive tags and applies priority/exclusion checks client-side.
• priority_tags results are fetched first, merged without duplicates, then capped.
• min_results permits one retry without the configured type restriction while retaining the same route filters.
• Reflect uses the same route filters; unsupported routed reflect or an explicit no-information response falls back to formatted routed recall.
• retain_tags are merged after global retain tags and before per-call tags.
• Low-signal suppression is opt-in and applies to both automatic prefetch and explicit recall; configured signal keywords bypass it.

Scope

This refresh is rebuilt from current origin/main and contains only:

• plugins/memory/hindsight/__init__.py
• tests/plugins/memory/test_hindsight_provider.py

It intentionally excludes Task Scope work, host-specific configuration/data, gateway changes, deployment scripts, and other product changes.

Tests

• HERMES_HOME=<temp> python -m pytest tests/plugins/memory/test_hindsight_provider.py -q -o 'addopts=' — 137 passed
• python -m ruff check plugins/memory/hindsight/__init__.py tests/plugins/memory/test_hindsight_provider.py — passed
• python -m py_compile plugins/memory/hindsight/__init__.py tests/plugins/memory/test_hindsight_provider.py — passed
• git diff --check origin/main...HEAD — passed

Coverage includes configuration parsing/schema exposure, signal precedence and config-order ties, no-match global fallback, prefetch routing before query truncation, positive/negative filters, fallback filtering, priority ordering, caps, bounded type broadening, reflect fallback, retain tagging, low-signal suppression, and safe service failure.

Risk

Mostly opt-in: route, retain, priority, fallback, cap, and low-signal behavior activate only when configured. The non-route reflect path now also passes the existing recall_max_tokens value to Hindsight so reflect and recall share the configured token ceiling.