feat(dream): add bd dream subcommand for memory consolidation (#3263)

[boaz-hwang]

Closes #3263.

Summary

Adds bd dream — a subcommand that consolidates the memory store (bd remember / bd memories) by asking an LLM to identify duplicates, stale references, and low-signal entries, then applies a structured forget / merge / update plan via the existing storage layer. See docs/DREAM.md (added in this PR) for the full design.

CLI surface:

bd dream [run]                   # default = run
bd dream run --dry-run           # show plan only
bd dream run --force             # bypass the time/churn threshold
bd dream run --check             # exit 0 if eligible, 1 if not (cron / hooks)
bd dream status                  # last run + threshold state

--json is supported on every subcommand.

What's included

• internal/dream/ (consolidator + prompt + safety guard + types) — all unit-tested with a messagesNewer fake (no network in tests).
• cmd/bd/dream.go (cobra commands + threshold logic + state read/write + apply pipeline) — integration-tested against embedded Dolt.
• cmd/bd/kv.go — adds \"dream.\" to the reserved-prefix list so users can't shadow dream's state via bd kv set.
• docs/DREAM.md + CHANGELOG.md (Unreleased entry).

Conventions followed

• Cobra parent-child pattern (hooksCmd / kvCmd style) with GroupID: \"setup\".
• CheckReadonly(\"dream\") on mutating ops, ensureDirectMode on every DB-touching op.
• store.GetAllConfig filtered by kvPrefix + memoryPrefix (inlined as in cmd/bd/memory.go:134-141).
• store.SetConfig / DeleteConfig followed by a single CommitPending(ctx, getActor()) per command run.
• --json flag respected in every subcommand (matches bd memories / bd recall).
• internal/compact/haiku.go retry / API-key / model resolution pattern reused for the SDK call.
• //go:build cgo on the embedded-Dolt integration test file.

Threshold semantics

• dream.min-interval-hours (default 24) — minimum hours between runs.
• dream.min-churn (default 5) — minimum memory-count delta since last run.
• 2-memory floor applies even with --force (nothing to consolidate otherwise).
• First run is always eligible.

State (dream.last-run-at, dream.last-run-status, dream.last-run-summary, dream.memory-count-at-last-run, dream.min-interval-hours, dream.min-churn) lives in the Dolt config table and syncs via bd dolt push like every other dream / sync-related key.

Test plan

• go test ./internal/dream/... — 8 unit tests pass (safety guard, malformed responses, eligibility logic).
• go test ./cmd/bd/ -run TestDream|TestKVReservedDream — 6 integration tests pass against embedded Dolt (status / status JSON / --check exit codes / churn gate / missing-API-key path / reserved-prefix guard).
• gofmt -l clean, golangci-lint run --build-tags=gms_pure_go 0 issues on changed files.
• make build clean.
• End-to-end manual: applied a real consolidation pass on a project with 12 memories — merged 2 newsletter-related entries into one richer entry. Safety guard correctly rejected an earlier run that would have deleted >50%.

Out of scope

• A pluggable internal/ai.Executor (sdk + claude-cli) so users on a Claude Code subscription can use bd dream (and eventually bd compact) without a separate API key. Follow-up PR is already prepared on a stacked branch and will be opened once this lands. Keeping it separate per CONTRIBUTING.md's "one issue per PR" rule.
• A bd dream config interactive wizard.
• A native scheduler / bd dream schedule helper.
• OpenAI / Ollama providers.
• Refactoring bd compact and bd find-duplicates --method ai onto the executor abstraction (lands with the executor PR).

Local verification recipe

export ANTHROPIC_API_KEY=sk-ant-...
cd <a project with several bd memories>
bd dream run --dry-run --force
bd dream run --force
bd dream status

🤖 Generated with Claude Code