feat: add session-based context window management

[Siddhant-K-code]

Stateful context windows with token budgets for long-running agent sessions. Entries are deduplicated on push, compressed through hierarchical levels as they age, and evicted when the budget is exceeded. Lowest-importance entries are compressed/evicted first.

Compression levels: full text → summary (~20%) → sentence → keywords. The preserve_recent setting keeps the N most recent entries at full fidelity.

Closes #31

Added

• pkg/session — SQLite-backed session store with 13 tests
• CLI — distill session create/push/context/delete
• HTTP API — /v1/session/create, /push, /context, /delete, /get (opt-in via --session flag)
• MCP tools — create_session, push_session, session_context, delete_session (opt-in via --session flag)

Docs

• README — Session Management section with CLI/API/MCP examples, updated architecture diagram, roadmap, config example, Pipeline Modules
• CHANGELOG — [Unreleased] section with session management
• FAQ — 4 new Q&As (What is Context Memory, What are Sessions, Memory vs Sessions, MCP tools list), updated roadmap status
• mcp/README — 8 new tool docs (memory + session), updated config examples, 2 new integration patterns
• examples/session_api.sh — Full API usage example
• mcp/claude_desktop_config_full.example.json — Config with all features enabled

Usage

# API server
distill api --port 8080 --session

# MCP server
distill mcp --session

# CLI
distill session create --session-id task-42 --max-tokens 128000
distill session push --session-id task-42 --role tool --content "file contents..." --importance 0.8
distill session context --session-id task-42

Review Findings Addressed

All 10 findings from code review resolved:

• P2: isDuplicate TODO, Create timestamp fix, enforceBudget loop
• P3: CompressionSavings rename
• Nits: removed unused ErrEmptyContent/CompressAge, sort.Slice, merged store constructors, word-boundary truncation

[Siddhant-K-code]

Code Review - PR #38

CI green (build + lint). 13 tests pass. 0 lint issues.

Verdict: Approve with fixes below

No blocking issues. Architecture follows the memory store patterns well.

Findings

Priority	File	Finding
P2	sqlite.go:437	isDuplicate full table scan per push entry. Add TODO noting O(n) scaling limit.
P2	sqlite.go:127-131	Create returns time.Now() twice instead of reusing the formatted timestamp. Subtle inconsistency with DB.
P2	sqlite.go:488-510	enforceBudget only compresses one level per push. A single large push can leave session over budget. Loop until within budget.
P3	sqlite.go:300-305	DedupSavings stat is misleading - rename to compression_savings.
Nit	session.go:14-17	ErrEmptyContent declared but never used. Remove.
Nit	session.go:148-152	CompressAge in Config never read. Remove or implement.
Nit	sqlite.go:632-645	sortCandidates uses insertion sort. Use sort.Slice.
Nit	session.go:193-225	sessionStoreFromFlags and sessionStoreFromConfig are nearly identical. Merge.
Nit	sqlite.go:575-583	LevelSentence truncation at 50 chars can cut mid-word. Truncate at word boundary.
Nit	session.go:224-225	Trailing blank line.

What's Good

• Scan-then-close pattern used consistently (learned from memory store review)
• preserve_recent keeps recent entries at full fidelity
• Importance-based eviction ordering
• handleContext accepts both GET and POST
• handleDelete accepts both DELETE and POST
• 13 tests covering all key paths

[Siddhant-K-code]

All 10 review findings addressed in 6302608:

Finding	Fix
P2 isDuplicate full scan	Added TODO comment
P2 Create timestamp inconsistency	Reuse time.Time instead of calling time.Now() twice
P2 enforceBudget single-pass	Loop until within budget or no progress
P3 DedupSavings misleading	Renamed to CompressionSavings
Nit ErrEmptyContent unused	Removed
Nit CompressAge unused	Removed from Config
Nit Insertion sort	Replaced with sort.Slice
Nit Duplicate store constructors	Merged into newSessionStore
Nit Mid-word truncation	Truncate at word boundary
Nit Trailing blank line	Removed

All 13 tests pass, 0 lint issues.