fix(store): reject empty observation titles before persistence

[tommanzur]

Summary

Closes #459.

Empty (or whitespace-only) observation titles were accepted by mem_save and
store.AddObservation, persisted locally, and queued for cloud sync. The cloud
server rejects them via ValidateSyncMutationPayload, blocking the entire
mutation queue silently — the caller receives success at write time and has no
feedback to retry with a valid title.

This PR pushes the same validation up to the write path, so the failure
surfaces immediately and can be corrected.

Changes

• internal/store/store.go: AddObservation rejects empty/whitespace titles
  with a clear error, before any DB access. Validated after stripPrivateTags
  so it matches exactly the value that gets persisted and later pushed to the
  cloud server.
• internal/mcp/mcp.go: handleSave returns an agent-actionable
  ToolResultError before reaching the store, with an example title format so
  the model knows what to provide on retry.
• Tests added for both layers (empty / whitespace / missing title).

Why this approach

Per the suggestion on #459, reusing the rule already enforced by
ValidateSyncMutationPayload (diagnostic.go) keeps validation in one logical
place. The fix is defensive at two layers — MCP for UX, store as the single
source of truth — so the CLI, MCP, and any future entrypoint inherit it.

Test plan

• go test ./internal/store/... — new tests for empty/whitespace titles + a positive case
• go test ./internal/mcp/... — handler returns ToolResultError, nothing persisted
• go build ./cmd/engram — binary builds clean
• Existing tests pass (no regression in dedupe, topic_key upsert, normalization)

Notes

• This is technically a behavior change for callers passing an empty title
  today, but those cases are already broken downstream (cloud sync blocks them).
  The change turns a silent failure into a loud, immediate one.
• stripPrivateTags replaces <private>…</private> with [REDACTED], not an
  empty string, so a title made only of private tags is not rejected — that
  is intended.

Out of scope

• Backfilling existing observations with empty titles in users' local DBs. The
  repair tooling and the SQLite workaround documented in Reject empty observation titles before persistence #459's comments handle
  that path. This PR prevents new occurrences.
• Title length / format constraints beyond non-empty. Can be a follow-up.

[tommanzur]

Thanks @ricardoarz-dev for the pointer to ValidateSyncMutationPayload — this PR reuses exactly that rule at write time, in store.AddObservation (single source of truth) plus an early agent-actionable error in handleSave. cc @jorocoimbra since you confirmed the repro and the stuck-queue recovery path.

One note for a maintainer: I couldn't add the type:bug label myself (external contributor — GitHub blocks labeling), so it'll need to be applied on triage.

Happy to adjust the approach or split anything further if it helps review.

[Alan-TheGentleman]

#459 is now approved and this PR has the type:bug label. The problem is valid: empty titles should fail before persistence instead of poisoning sync later. Next step is code review after required checks are green.