fix(quick-router): accept a leading please on forget commands

[galuis116]

Summary

Closes #921. memory_forget_query matches its prefixes at token 0 and already
stripped a trailing please, but never a leading one — so please forget the wifi password matched no prefix.

Unlike the sibling leading-please gaps, missing here does not abstain.
route runs memory_forget_query and then memory_recall_query, whose
household / note / semantic matchers are contains-based. The utterance was
claimed by recall, and the router emitted a confidently wrong tool:

"please forget the wifi password"
    -> memory_recall{query: "please forget the wifi password", limit: 3}
expected
    -> memory_forget{query: "wifi password"}

utterance	before	after
Please forget the wifi password	memory_recall{"please forget the wifi password"}	memory_forget{"wifi password"}
Please forget my old locker combination	recall / ABSTAIN	memory_forget{"old locker combination"}
Please delete what you know about my car	ABSTAIN	memory_forget{"my car"}

Why this one is worse than the sibling gaps

A missed route costs a fallthrough to the LLM, which can still recover. Here
the user asked to delete a memory and the router reads it back instead:

• the secret is spoken aloud rather than erased,
• the query is please forget the wifi password, which is not a stored fact
  and matches nothing useful,
• the user gets a plausible-looking answer and no signal that the delete never
  happened.

A privacy-intent utterance silently becoming a retrieval is a correctness bug
with a safety edge, not just a coverage gap.

Changes

• Strip a single leading please before the prefix loop in
  memory_forget_query, mirroring scene_or_routine_activation_request and
  play_media_request.
• The bare-pronoun abstention (that / it / this) and the
  trailing-please trim inside the loop are untouched, so please forget about it still abstains rather than deleting the literal word.
• New regression test forget_command_accepts_a_leading_please.

Real Behavior Proof

• I have built and run the affected code locally (or noted why I could not).
• I have verified the change end-to-end on Jetson hardware.
• I have NOT verified on Jetson hardware, and I explain the equivalent verification path or validation gap below.

Tested profile / hardware (check all that apply):

• jetson
• raspberry_pi
• portable_sbc
• laptop
• mac
• CI-only / docs-only
• Not run locally

What I ran

x86_64 Linux laptop (Ubuntu 22.04 LTS, kernel 6.8.0-136-generic,
rustc 1.96.0 / cargo 1.96.0). No Jetson available to me — see the validation
gap below.

cargo test -p genie-core
cargo clippy -p genie-core --all-targets
cargo fmt -p genie-core -- --check

Plus a temporary probe binding route() directly, so I could read the actual
emitted ToolCall on both main and this branch — the whole point of this bug
is which tool fires, so an is_none()-style assertion would not have shown
it. The probe was removed before committing; only the permanent regression test
remains.

What I observed

Probe output on main — the delete request answered as a lookup:

please forget the wifi password -> memory_recall {"limit":3,"query":"please forget the wifi password"}
forget the wifi password        -> memory_forget {"query":"wifi password"}

Probe output on this branch:

please forget the wifi password          -> memory_forget {"query":"wifi password"}
please forget my old locker combination  -> memory_forget {"query":"old locker combination"}
please delete what you know about my car -> memory_forget {"query":"my car"}
please forget about it                   -> ABSTAIN

The polite call is now byte-identical to the bare call, and the bare-pronoun
abstention still holds under a leading please.

cargo test -p genie-core: 1005 lib tests passed, 0 failed, 6 ignored, and
all integration suites green. cargo clippy --all-targets produced no
warnings; cargo fmt --check produced no diff.

Validation gap

I could not verify on Jetson hardware. The equivalent verification path: this
change is confined to crates/genie-core/src/tools/quick.rs, a pure
string→ToolCall function with no I/O, no async, no hardware dependency, and
no model involvement. Its entire contract is which ToolCall comes out for a
given utterance, asserted directly above. The emitted call is byte-identical to
the one the bare utterance already produces and already dispatches on Jetson
today, so the memory subsystem downstream sees nothing new. Nothing in the diff
touches audio, the dashboard, or the model prompt.

Test plan

1. Cherry-pick forget_command_accepts_a_leading_please onto main alone and
   run it — it fails, and fails loudly: main returns memory_recall, not
   None.
2. Check out this branch, rerun — passes.
3. On a device with memories stored, say "remember my wifi password is X", then
   "please forget the wifi password", then "what's my wifi password" — confirm
   the fact is gone rather than read back.

Notes for reviewers

• The test asserts the polite call equals the bare call by name and
  arguments, not merely that some route exists — a partial fix that stripped
  please from the tool selection but not from the query would still fail.
• No prompt growth, no new dependencies. The 4096-token Jetson context contract
  is untouched.
• Behavior is unchanged for every utterance that does not begin with please .

[coderabbitai]

📝 Walkthrough

Walkthrough

The deterministic router now handles leading “please” in memory forget and delete commands. Tests verify equivalent query extraction for polite and bare forms while preserving abstention for bare-pronoun references.

Changes

Memory Forget Routing

Layer / File(s)	Summary
Forget prefix handling and regression coverage crates/genie-core/src/tools/quick.rs	Leading "please " is stripped before forget/delete prefix matching, with tests covering routing, query extraction, trailing politeness, and pronoun-based abstention.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• GeniePod/genie-claw#850: Updates the same router’s leading-please handling for shopping-list commands.
• GeniePod/genie-claw#892: Applies the same leading-please stripping pattern to media requests.
• GeniePod/genie-claw#895: Applies the same deterministic routing adjustment to scene or routine activation.

Suggested reviewers: matedev01, ultrahighsuper, yurii214

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The change matches issue #921 by routing leading-please forget/delete commands to memory_forget while preserving pronoun abstention and adding regression tests.
Out of Scope Changes check	✅ Passed	The PR appears tightly scoped to the quick-router fix and related tests, with no obvious unrelated changes.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: quick-router now accepts a leading “please” on forget commands.

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Fix failing CI checks

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}