Roadmap: known gaps and deferred work

[HackPoint]

The index of everything measured but not done, and of the claims that were checked while building
it. Each linked issue states what is wrong, the measured evidence, acceptance criteria, and an exact
command to test it.

Numbers come from cargo test --release -p lumen-mcp --test efficiency -- --nocapture, which prints
them and asserts them, so they cannot rot into stale prose the way the last set did.

Before a release — milestone 1.6.0

• Release 1.6.0 #6 — Release 1.6.0. A minor bump: lumen doctor, lumen show and the degraded banner are
  new features.
• Daemon reconnect noise rotates the tray diagnostics out of the log #7 — Daemon reconnect noise rotates the tray diagnostics out of the log. 434 KB/hour
  against a 256 KB cap with KeepOne = the log rotates every ~35 minutes, taking the evidence
  release logging was just added to preserve.

Coverage backlog — 653k tokens read whole, measured

Milestone Backlog. None of this is a routing
failure; it is file kinds the optimizer does not cover.

• Intercept .js and .jsx — no new grammar needed #8 — .js / .jsx, 109,914 tokens. The cheapest item here: tree-sitter-typescript is
  already a dependency and its grammar is a superset of JavaScript, so no new grammar is needed.
• Outline Markdown by its headings #9 — Markdown by headings, 250,403 tokens over 64 files. The largest single opportunity.
  Needs a line scanner rather than tree-sitter, with fenced-code state tracked.
• .out is intercepted but .output is not #10 — .out is intercepted but .output is not, 9,547 tokens. A one-word gap in the
  extension list.
• Decide whether .scss and .css are worth intercepting #11 — .scss / .css: decide, do not assume. 180,448 + 63,175 tokens looks large until
  you see that 83% of the scss total is one 3,116-line file read 11 times on one machine.
  Wants a second corpus before committing.

Economics and hygiene

• Files just over the 300-line threshold fall below the interception break-even #12 — 17 of 31 corpus files fall below the 5,383-token break-even. Raising the threshold
  is the obvious move and is not clearly right; wants a measured sweep, not a guess.
• The sessions table is declared in the schema and never used #13 — The sessions table is declared and never used. 0 rows while turns holds 66
  distinct session ids. Populate it or drop it. (is_subagent has the same shape — 0 across all
  4,446 rows — so resolve both together.)

Considered and rejected — kept so the reasoning is searchable

Closed as not planned. Reopening one is how the decision gets revisited.

• Decision: not intercepting .yml #14 — .yml. A partial YAML read is not valid YAML, so an outline invites a confidently wrong
  answer, and 31,010 tokens is below break-even anyway.
• Decision: not intercepting .json #15 — .json. The largest on record is 271 lines against a 300-line threshold, so
  interception can never fire. Structurally out of reach.
• Decision: leaving 3.4% of the reported saving with unverified provenance #16 — The 3.4% of the saving with unverified provenance. Rows predating token_source. The
  ledger is append-only and the count is already disclosed; rewriting them would trade a disclosed
  unknown for an undisclosed guess.
• Decision: Linux compilation cannot be verified on the development machine #17 — Linux compilation cannot be verified locally. No target, and the Tauri crate needs GTK.
  This is how a RunEvent::Reopen match arm reached CI red. The mitigation is procedural.

Checked while writing this, and resolved

Three claims went into the CHANGELOG and README without a test behind them. All three are now
settled, and two of them turned up a further defect on the way.

• RunEvent::Reopen was asserted to work but never verified #18 — RunEvent::Reopen was asserted to reveal the window and never verified. It does —
  AppKit delivers reopen under ActivationPolicy::Accessory. Only testable from a bundle.
• The hidden-icon preference fix was unverified in the real app bundle #19 — The hidden-icon preference fix was verified only against the Lumen defaults domain,
  not the io.speedata.lumen domain the bundle owns. It works — and verifying it exposed a second
  key form, NSStatusItem VisibleCC <name>, that the fix was ignoring entirely.
• lumen doctor reported a menu-bar manager that was never installed #20 — lumen doctor reported a menu-bar manager that has never been installed. It matched
  names as substrings of the whole ps output, and "ice" is a substring of "Services": 130
  processes matched on a stock install. It would have invented a phantom cause on almost any Mac.

Two more found the same way, fixed in 8f998b1: a test that opened the developer's real browser on
every cargo test --workspace, and a published metric — "0 of 1,573 eligible reads leaked" —
asserted as an invariant it never was, since the intercept's own fail-open guards produce exactly
that shape by design.

The pattern in all five: they were found by running the thing and reading the real output, not by
reasoning about the code.

[HackPoint]

On the numbers in this issue. Every figure above is a snapshot taken 2026-07-31, not a constant. The ledger grows as the tool is used and the corpus changes with every commit that adds or crosses a 300-line file — the break-even count already moved from 17 to 16 within an hour of being filed (see #12).

Re-measure rather than trusting the text:

cargo test --release -p lumen-mcp --test efficiency -- --nocapture

That prints the coverage table, the backlog per file kind, the break-even shortfall and the priced ledger, and asserts the bounds. If a number here disagrees with that output, the output is right.