feat: tag management CLI and trait extensions

[asvarnon]

Summary

Add tag management CLI commands and trait extensions for user-defined and auto-generated entry tags. Enables filtered queries, organized context retrieval, and integration with importance detection category labels.

Note: Tag tables (tags, entry_tags, indexes) are created in #12 (Schema v2 migration). This issue covers the CLI surface, trait extensions, and auto-tagging logic only.

Motivation

Entries currently have kind (Manual, PreCompact, Auto) but no user-defined categorization. As the store grows, users need to:

• Filter context by project, language, or topic
• Preserve important entries from eviction
• Scope SessionStart injection to relevant domains

Importance detection (#49) will auto-generate category tags (corrective, reinforcing, stateful, decisive) — this issue's tag infrastructure supports that.

Scope

In Scope

CLI Surface:

cf save --kind auto --tags "rust,security"      # manual tags at save time
cf query --tags "rust" --format text             # filter to tagged entries
cf query --tags "rust,auth" --tag-match all      # AND logic (default)
cf query --tags "rust,auth" --tag-match any      # OR logic
cf tags list                                     # all tags with entry counts
cf tags rename <old> <new>                       # rename a tag
cf tags delete <tag>                             # remove tag (entries kept)
cf tags merge <tag1> <tag2>                      # combine into one

Trait Extensions:

// ContextStorage additions
fn add_tag(&self, entry_id: &str, tag: &str) -> Result<()>;
fn remove_tag(&self, entry_id: &str, tag: &str) -> Result<()>;
fn get_tags(&self, entry_id: &str) -> Result<Vec<String>>;
fn list_all_tags(&self) -> Result<Vec<TagInfo>>;
fn rename_tag(&self, old: &str, new: &str) -> Result<usize>;

// Searcher addition
fn search_with_tags(&self, query: &str, tags: &[&str], match_all: bool, limit: usize)
    -> Result<Vec<ScoredEntry>>;

Auto-Tagging (Phased):

• Phase 1 — Metadata-based: auto-tag by entry kind, size-based large for entries >8,000 tokens
• Phase 2 — TF-IDF keyword extraction: may be superseded by importance detection's recurrence scoring (feat: implement session frequency and recurrence scoring #47)
• Phase 3 — LLM-generated tags: extension-layer only, out of scope for core

Assembly Integration:

• Tag filtering applied before BM25 scoring (narrows candidate set)
• Optional score boost for tag-matched entries (configurable multiplier)
• New eviction policy: LruPreserveTagged — evict oldest untagged first

ContextEntry Extension:

pub struct ContextEntry {
    // ... existing fields ...
    pub tags: Vec<String>,  // populated on retrieval via JOIN
}

Out of Scope

• Tag table DDL (created in Schema v2 Migration: Unified Runtime Adapter + Importance Detection Foundation #12)
• Schema migration (handled in Schema v2 Migration: Unified Runtime Adapter + Importance Detection Foundation #12)
• Importance detection category auto-tagging logic (handled in feat: classify extracted passages into importance categories #49)

Acceptance Criteria

• cf save --tags persists tags on new entries
• cf query --tags filters results by tag (AND and OR modes)
• cf tags list/rename/delete/merge subcommands work correctly
• Trait extensions implemented on SqliteStorage
• Phase 1 auto-tagging: kind-based and size-based tags applied at save time
• Tag-filtered search integrates with BM25 scoring
• Unit tests for all trait methods
• Integration tests for CLI tag commands

Dependencies

• Schema v2 Migration: Unified Runtime Adapter + Importance Detection Foundation #12 — Schema v2 must land first (tag tables created there)

Complexity: Medium

Trait extensions, CLI plumbing, auto-tagging logic. No schema work needed.

[asvarnon]

Dependency #34 has been implemented and shipped in v0.3.0. This issue is now unblocked — BM25-targeted retrieval, config file, and FTS5 preprocessing are all in place as the foundation.

[asvarnon]

Scope note after the library refactor: the CLI half of this is obsolete (no CLI binary anymore), but the underlying need is still valid as a library feature — the tags/entry_tags schema survived the v3 migration but has no public API on the ContextForge facade yet. Keeping open as a library tag-management enhancement; the CLI framing no longer applies.