Compact citation marker scales O(N) and forces over-citation of all injected memories

[smoochy]

Summary

render_inline_compact_with_registry (in plugin/src/claude_smart/context_format.py) builds a single citation marker that concatenates every injected memory's linked title+URL and instructs the model to "copy this final marker exactly." This has two costs that grow with the number of injected memories (N):

1. Token scaling. The marker is O(N) — it embeds every item's title and URL, and each URL also already appears inline in the item's open: segment (duplicated). The normal (render_inline_with_registry) path instead gives an example marker and lets the model format only what it used.
2. Over-citation (correctness). Because the pre-built marker lists all N memories and the instruction says to copy it exactly, the model is nudged to cite memories it did not actually use — regardless of the WHEN_TO_CITE_COMPACT counterfactual gate above it.

This is deliberate (there is an explicit test: tests/test_context_format.py asserts "copy this final marker exactly with markdown links"), so this is a design-tradeoff report, not a bug claim — presumably the determinism helps citation reliability on hosts like the Codex TUI.

Evidence

Measured with the repo's real renderers (tiktoken cl100k_base), synthetic playbooks/preferences at defaults (citations on, markdown links):

N	normal render	compact render	delta
5	693	582	compact −16% (smaller)
20	1797	2022	compact +12.5% (larger)
50	4005	4902	compact +22.4% (larger)

With CLAUDE_SMART_CITATIONS=off (marker removed) compact is consistently ~14% smaller at all N — i.e. the crossover is entirely the citation marker, and past ~5 injected memories the "compact" path is larger than the normal one it is meant to compress.

(Full write-up + reproducible scripts in a downstream analysis; happy to share.)

Proposed direction

Mirror the normal path in the compact citation instruction: provide an example marker + the | separator guidance (as citation_instruction already does) instead of a pre-concatenated all-N "copy exactly" marker. The model then cites only the memories it used, the marker is O(used) not O(N), and URLs are not duplicated. The WHEN_TO_CITE_COMPACT gate stays.

If the determinism is intentional for a specific host, an alternative is to cap the pre-built marker (e.g. only when N ≤ small threshold) and fall back to example-guidance beyond it.

Glad to open a PR if this direction is welcome.