fix both criterion 6 failures at the cause

[ayushcodes10]

The saves bar and the bad-merge bar were both failing for reasons in the code
rather than in the trial.

The bad merges were still live

A confirmed resolved_to does two things, and the second is easy to miss: it
points the episode's facts at the named node and appends the mention to that
node's aliases. That alias is what makes a misdirected resolution a merge
rather than a misfiled fact, and it is criterion 6's third bar word for word -
two distinct entities in one node.

The 2026-09-02 cleanup deleted the misdirected edges and left the aliases. For
eight days node_embedding table answered to Eigon billing profile and AGE graphid column type to Eigon GST tax rule. _exact_match checks aliases, so
this was not historical damage: the next mention of either name was one retired
node away from landing on an embedding table's lesson.

echo-memory unmerge is the missing half of that cleanup - it lists every node
carrying another node's name and refuses to act on its own, because a genuine
duplicate pair on its way to being merged properly looks identical from here.
Both scopes of the author's store are now clean.

Running it found a third merge nobody had seen, made the same day by
session c87a143c-billing-hardening: Eigon warm-base ALB sharing absorbed
into Eigon API tokens page. Recorded as an observation rather than quietly
repaired, so the trial count is now 3, not 2.

The scope check from #44 would not have caught any of them

Every guessed id belonged to the caller, in a different project, so ownership
held. Owning a node is not the same as it being the right node.

resolve_entities now also asks whether it would ever have offered this node
as a candidate for this mention - candidates are drawn above low_threshold,
so an id pointing below it cannot be an answer to a question this server asked.
Measured before picking the bar rather than assumed:

pair	similarity
Eigon billing profile / node_embedding table	0.030
Eigon GST tax rule / AGE graphid column type	0.041
Eigon warm-base ALB sharing / Eigon API tokens page (today's)	0.259
AGE / Apache AGE - hardest legitimate confirmation on record	0.497

All three incidents fall below 0.45; the legitimate confirmation clears it. An
exact name or alias match skips the check, since there the id is redundant
rather than doubtful.

The saves bar could never have been met

It required the entire store to have attributed facts before any cross-tool
save could count. The danger is real - 'unknown' is unequal to every real agent
id, so a save citing an authorless fact satisfies written_by != recalled_by
because one side is missing, not because two tools were involved - but the
scope was wrong. It refused two saves naming real tools at both ends because of
twenty-seven unrelated facts in sessions holding nothing attributable, which
the check's own message called unrecoverable. An unclearable bar is not a gate.

The exclusion is now per save, against that save's own two agent ids.

reattribute --agent recovers the authors that are recoverable, and only
those: a session is one tool's conversation, so a session with both attributed
and unattributed facts has already said who wrote the rest. Three of the thirty
came back that way. The other twenty-seven sit in sessions with no attributed
fact at all and stay unattributed - a session claimed by two agent ids is
refused too, because taking the more frequent one is a guess wearing a majority
vote.

Where the trial actually stands

Nothing here manufactures evidence. After the fixes: 2/3 cross-tool saves
(a third has to be earned by a real one), 1 duplicate (within budget),
3 bad merges (one more than before, because the repair tool found one).

515 tests pass. Every new test was checked against the unfixed code first.

🤖 Generated with Claude Code

https://claude.ai/code/session_0191FunuahgEA734jXUrxzJm