Every letter belongs to the graph

[rvaccone]

A verb sitting on a letter costs an address forever, and an address is the
scarcer thing. Three verbs were doing it. All three move off, and
Gestures.reservedLetters is now empty — the whole alphabet belongs to the
graph, and is meant to stay that way.

The key map

verb	was	now	why
attention timeline	lode X	retired	least-used verb in the instrumentation, by more than two orders of magnitude against the address it blocked
layout undo / redo	lode Z / ⇧Z	lode ← / lode →	Layout.swift already calls it "one global timeline"; walking one is what arrows mean
flip orientation	lode O	lode \	already moved in 0.14.2 — the roster just never heard

Undo becomes a two-key pair like [ / ] rather than needing shift, and
costs no punctuation: `, -, and = all stay free. Since lode is
right-command it only shadows right-⌘←, which is not how anyone types
⌘←.

Retiring the timeline takes FocusHistory with it — it had no other caller,
and it was recording every focus change in the process.

One definition of the reserved set

It had been restated in six places, and each of the last three key moves
left copies behind:

	before	after
Gestures.reservedLetters	—	[] (the only definition)
Gestures.graphLetters	filtered o, x, z	derives from the above
Gestures.roster flip-orientation	keys: ["o"]	keys: ["\\"]
Gestures.roster layout-undo	keys: ["z"]	keys: ["left", "right"]
Config.reservedTopLevel	["z", "x"]	reads Gestures
Advisor.reservedLetters	["x", "z"]	reads Gestures
cheat sheet GuideRow	key: "X"	(row removed)

Bugs this fixes

1. The flip-orientation toggle was wired to the wrong key. 0.14.2 moved
   the verb from O to \ in the engine but not in the roster. Because the
   engine disables by key, flip-orientation: false disabled lode O —
   breaking an unrelated graph address — while lode \ kept flipping. The
   verb could not be turned off, and turning it off broke something else.

2. The cheat sheet taught a key the engine did not dispatch. The verb list
   in cheatSections() is hand-written, so it drifts silently.

3. mnemonicLetters dropped every parenthesised word. It split on spaces
   and took each word's first character, so "(xonar)" contributed nothing.
   Lodestar generates these names itself (Graph.Target.label renders a
   profile as "Brave (Xonar)"), so the distinguishing half of every
   browser-profile name was invisible and no profile could ever be offered a
   shorter address. "1Password" was hidden behind its digit the same way.

4. shorten could not clear its own probability gate. It priced
   uncertainty as perUseSavedSD = residualSD * saved — a fixed multiple of
   the saving — so the signal-to-noise ratio was always 1 / residualSD.
   With learningCost == 0 the Monte Carlo condition collapses to
   saved > 0, whose probability is Φ(1 / residualSD): a constant,
   independent of the size of the saving and of how often the chain is
   used.

   saving=0.05s/use -> p=0.8497
   saving= 0.2s/use -> p=0.8497
   saving= 2.0s/use -> p=0.8497   (identical)
   

   At residualSD = 0.966 that is 0.850 against a 0.9 gate, so no
   shorten could ever be offered, at any usage. The saving is a difference
   of two means, so it is now priced by how well those means are known: the
   chain being replaced tightens as √n, while the proposed chain keeps the
   full population spread because it has never been typed. bindCandidates
   already did this correctly via record.commit.se.

5. The Monte Carlo was not replayable, despite saying it was.
   Advisor.Random is documented as "Deterministic LCG so the Monte Carlo is
   replayable in tests" but was seeded from UInt64(abs(key.hashValue)), and
   Swift randomises hashValue per process — three runs of "b x".hashValue
   give three different values, so every run drew a different sample and any
   candidate near p = 0.9 would appear and vanish between runs. Now seeded
   from a stable FNV-1a hash, which is also total where abs(Int.min) would
   have trapped.

6. The test that should have caught bug 1 could not.
   testKeysAreDisjointAndCoverTheIdleDispatch claimed to check "every
   idle-state key the engine dispatches" but hardcoded the expected set, so it
   drifted in lockstep with the roster and actively asserted the stale mapping
   was correct. It now asks the engine — pressing every key in Keys.ansi
   and treating anything that is not .passThrough as claimed — and compares
   that against the roster.

Label is not target

Caught by review, and a regression this change introduced.
shortenCandidates built its edit from leaf.label — for a browser
profile, the rendered "Brave (Xonar)" rather than the bindable
"brave:xonar". It was unreachable while mnemonicLetters could only ever
yield b for a profile leaf; fixing the mnemonics is precisely what made
it reachable. Accepting the headline offer would have gone
AppIndex.entry(named: "Brave (Xonar)") → no exact match → Fuzzy.rank →
plain Brave, profile dropped in silence.

• GraphTarget.configValue sits beside label, so the two strings are
  named and documented rather than conflated.
• Advisor.Leaf replaces the (chain:label:) tuple and carries both.
• ConfigEdit.bindApp → .bindTarget, because the edit holds what a config
  line writes.
• Recommendation.display (optional) lets the chip show
  lode X → Brave (Xonar) while committing brave:xonar — a chip reading
  lode X → brave:xonar would quote the machinery at someone who only asked
  for their browser.
• addTargetToGraph binds a profile reference as written and checks the
  profiles registry, rather than sending it through the app index.

shorten had no test at all before this. It does now, and it fails
against the old behaviour — verified by reintroducing the bug.

Result

The first actionable recommendation the coach has ever produced on real
data:

worth looking at
  lode B X fires 547×/week · it has earned lode X   (≈116s/week, p 1.00)
  lode V Z fires 44×/week · it has earned lode Z    (≈18s/week, p 1.00)

Both carry a ConfigEdit, so they are offerable rather than report-only, and
the first clears the coach's 30s/week debut floor by nearly 4×. Reproducible
across consecutive runs.

Testing

swift test — 518 tests, 0 failures. Six FocusHistory tests retired with
the feature; three added (the profile-shorten regression, the engine-derived
dispatch check, and the whole-alphabet check). Updated every case that
encoded the old reserved set — GesturesTests, StarterGraphTests,
EngineTests, AnalysisTests, ConfigBuildTests, ModifierTapTests.

Config.parse still refuses a graph binding on a reserved letter. With an
empty set it has nothing to refuse, which is deliberate: it stays as a guard
in case a verb ever moves back onto a letter, and its message no longer names
particular letters.

Pre-0.17 configs carry a gestures.back-forward: switch for a verb that no
longer exists; ConfigDefaults.normalized absorbs it the same way it
absorbed the searcher → launcher rename, so old files do not start
reporting an unknown key.