fix(dreaming): version the topic map so an extractor fix reaches existing workspaces (v0.281.2)

[vikasprogrammer]

Caught while verifying #500 on the live fleet: the extractor fix had no effect on any workspace that was already running.

state.topics is a cumulative map — counts compound across passes and decay only on a 21-day half-life. Fixing what gets extracted doesn't remove what was already extracted. The live instapods state:

stored topics: 300 | top: claude(62), drafts(61), sweep(59), automated(58),
                          incremental(58), support(58), phases(58), every(55)

Every one of those is an artefact of a single shouted prompt header (AUTOMATED INCREMENTAL SUPPORT SWEEP …) that #500 stopped extracting — and every one would have kept headlining the guidance line in every agent's system prompt for weeks, being reinforced faster than it decayed.

Fix

DreamState gains topicsVersion. When a stored state's version is older than TOPICS_VERSION, its topic map is cleared and rebuilt from the current corpus.

The reset runs before the no-activity early return — otherwise a quiet workspace would keep serving the stale line until its next busy pass. Audited as learning.topics.reset with the dropped count.

The alternative was hand-editing five production databases; this way every tenant (including insta-ai and expresstech, which nobody is looking at today) self-heals on its next pass, and the next extractor change is a one-line version bump.

Verified

Against a copy of the live instapods state:

BEFORE: topics=300 | version=undefined
  top: claude(62), drafts(61), sweep(59), automated(58), incremental(58), support(58)
AFTER : topics=0   | version=2
  guidance: (no topic line — map rebuilding)

The topic line disappearing until the map rebuilds is the intended behaviour: saying nothing is correct, saying sweep, automated, drafts was not.

137/137 governance + 18/18 tier-A policy.

🤖 Generated with Claude Code