fix(memory): stop a turn recalling the question it is answering

[sulthannauval]

Problem

Auto-save writes each user message to memory before recall runs, so by the
time the [Memory context] block is built the store already holds a verbatim
copy of the question being asked.

That copy is worthless as context — the same text sits in the prompt directly
below the block — and it is actively harmful. Being a perfect lexical match it
takes the top rank, and scores are normalised relative to the best hit, so a
single self-echo pushes every real fact under min_relevance_score.

The effect, measured on a live store with the same query and the same config,
differing only in auto_save:

auto_save	keys reaching the prompt
true	["user_msg_ddbebf17-…"] — one entry, its own question
false	["long_note","release_cadence","daily_deploy","deploy_window","roadmap_note"]

With the default settings, curated memory was effectively never reaching the
model. Clearing the conversation category did not help — the next turn
re-writes its own entry before recalling.

This was invisible until #417 made injection visible.

Change

build_memory_context drops entries whose content is the message being
answered, then re-ranks the survivors.

Two details carry the fix, and each is separately mutation-tested:

• Dropped before the score threshold, not after. By the time the threshold
  runs, the echo has already defined the ranking everything else is measured
  against — filtering it later changes nothing.
• Re-ranked after the drop, so "relative to the best hit" means the best
  usable hit.

One place, three callers: the agent, the channel dispatcher and the CLI loop
share this builder.

An entry that merely mentions the topic is untouched — only a verbatim echo
is dropped, matched after collapsing whitespace so a re-wrapped copy still
compares equal.

Known limit

Asking the same question verbatim N times leaves N stored echoes. Past the
small over-fetch margin the usable page shrinks by one per extra echo — at four
identical turns the block went from five entries to four. Widening the recall on
every turn to buy headroom for verbatim repeats costs more than the case is
worth; what is lost is the weakest-ranked entries, and nothing incorrect enters
the prompt. Documented on the constant.

Not changed

recall_layered puts conversation-scoped hits ahead of shared memory, and both
tiers share one entry budget. That balance is a design decision, not a defect,
and it is left alone here.

Validation

• cargo test --lib memory:: / agent:: / channels:: and --test agent_e2e,
  --test memory_comparison — 1311 passed.
• Mutation-tested both halves: skipping the re-rank fails 2 tests; moving
  the drop after the threshold fails the same 2.
• Strict delta gate — clean.
• Live, auto_save = true, gateway and TUI:

↺ recalled 5 memories: long_note, release_cadence, daily_deploy, +2 more

Stable across repeated turns (2, 3, 4).

Risk / rollback

This changes what every turn sends the model — that is the point, but it is
the blast radius. One file, one function, no config or schema change. Revert the
commit.

[github-actions]

PR intake checks found warnings (non-blocking)

Fast safe checks found advisory issues. CI lint/test/build gates still enforce merge quality.

• Missing required PR template sections: ## Summary, ## Validation Evidence (required), ## Security Impact (required), ## Privacy and Data Hygiene (required), ## Rollback Plan (required)
• Incomplete required PR template fields: summary problem, summary why it matters, summary what changed, validation commands, security risk/mitigation, privacy status, rollback plan

Action items:

1. Complete required PR template sections/fields.
2. Remove tabs, trailing whitespace, and merge conflict markers from added lines.
3. Re-run local checks before pushing:
   • ./scripts/ci/rust_quality_gate.sh
   • ./scripts/ci/rust_strict_delta_gate.sh
   • ./scripts/ci/docs_quality_gate.sh

Run logs: https://github.com/RantAI-dev/RantAIClaw/actions/runs/31090472574

Detected blocking line issues (sample):

• none

Detected advisory line issues (sample):

• none