Smiteless v0.9.72

• THE OUT: the game you're LOSING, and the fourteen minutes nobody counts

New feature, and it is the first one aimed at the half of the climb this app has never
touched. Every surface here so far attacks the same number: win the game you're in. None of
them have ever touched the DENOMINATOR - and a climb is LP per HOUR.

Half of that hour is spent inside games that were decided ten minutes ago. A 38-minute loss
and a 24-minute loss cost you exactly the same LP; the difference is fourteen minutes you
will never queue with. And the mirror of it costs even more: the game your team throws away
at the 15:00 vote that was never actually lost.

THE CLOSER owns the game you're winning and says nothing at all when you're behind - that
was always deliberate, and it left half the map dark. THE OUT is that other half, and it
answers one question at the only moment anybody asks it:

THE OUT     -4.2k · out: baron 0:35 · won back 3.7k of 7.9k

• An out is a fact with a clock on it, never a mood. Baron in range and actually
  contestable. Elder, or a soul point that's yours to take. Death timers long enough that
  one won fight is the whole map - "their deaths cost 52s", computed off their own levels
  on the live clock. A comp that out-scales theirs. Or a base they still haven't opened,
  which means they're up gold and have to come through you to spend it.
• Because the most expensive thing in a losing game is a team that stops playing one it
  could still win. From 15:00 - the first surrender window; there's no such thing as a
  "decided" game before one exists - and only while you're 2k+ down, which is the exact bar
  THE CLOSER calls ahead, mirrored. The two are one read of one map and can never both be
  talking.
• The number nobody has ever shown a losing team: what you've WON BACK. THE CLOSER
  tracks the peak of a lead because a lead being given back is a game being thrown in slow
  motion. This is the mirror - your worst deficit against where you are now, in the same
  measured gold - and it is the comeback happening before anyone on your team can feel it.
• CALL IT, and it is deliberately the hardest verdict in the app to reach. Four things
  have to be true at once, all measured off the game, none of them modelled: past 20:00, 8k
  down, an inhibitor of yours open (or a nexus turret gone), a 5v5 you lose by a mile - and
  no live objective out on the board. That isn't "you're losing", that's a game with no
  remaining mechanism. Then it says the true thing: the LP is already spent, and the
  minutes are not.
• It reaches you on the death screen too, where you're actually sitting when you decide
  whether to keep playing. HOW YOU WIN on the Death Brief is now this read, in the same
  words the widget is showing.
• It never speaks aloud, and it never votes for you. 100% read-only, like everything
  else in here.

The win% chip is gone, and good riddance. It was the one number on the widget that
changed no decision you were about to make, it was modelled rather than measured, and
"BEHIND ~34%" is the single most likely pixel in this app to make somebody stop playing a
game they could still win. In its place is the measured thing the guards beside it are
actually reasoning about - TEAM -4.2k, the same gold gap THE CLOSER calls a lead and THE
OUT calls a deficit - and THE OUT is what says the rest.

Tested, and the tests are mostly about what it must NOT do. 19 verdict fixtures, then
10,000 fuzzed game states asserting that CALL IT never fires without all four of its
facts, never hides in a quiet row, and never appears while a live objective out is on the
board - plus that every quiet row is genuinely one line so nothing gets clipped. Then the
number that actually matters: 600 simulated games, gold as a random walk with a per-game
drift so some teams recover, structures falling out of a sustained deficit the way they
really do - and 0% of the games it wrote off ever came back to even. That check is a
tripwire, not a trophy: loosen any bar and the retraction rate climbs, which is the only way
this verdict could ever rot without anybody noticing. Every bar was mutation-tested to prove
the suite fails when it's moved.

And one bug caught by the test that tests the wiring rather than the math. The first cut
read the enemy's fallen turrets when asking "how deep are they into you", so a team whose
own base was already open would have been told nothing of theirs was - the exact frame where
being wrong matters most. No fixture can catch that, so there's now an end-to-end check that
drives the whole guard off a real-shaped live payload, and it fails the build if that ever
comes back.

One brain, again. The power curve that decides "you out-scale them" in game is the same
table champ select graded the draft with, and the team gold gap is now one function shared
by THE CLOSER, THE OUT and the widget's chip - so nothing on your screen can quote a
different number than the guard sitting next to it.

Smiteless v0.9.71

• THE POOL: your champions, priced in your own LP - and the stat that finally stops lying to you

New feature, and it is aimed at the single biggest lever there is. Smiteless has enforced
champion-pool discipline since MAX ELO shipped. It has never once told you what your pool
should BE. Now it does - in the same LP your habits are priced in.

Last version priced the five habits your profile grades and named the one worth fixing.
The other half of the climb - and by every study, the bigger half - is which champions you
press Find Match on. The old answer to that was three win-rate bullets on your profile
("play more Sett 58%"). That's not an answer. 58% of what? Better than what? Worth what?

THE POOL - 9 champions over your last 57
  Sett            +38 LP / 10 on it     14W-6L over 20    70% on it vs 45% otherwise
  Ornn              +11pp lean           6W-3L over 9
  Darius          -44 LP / 10 on it      2W-8L over 10    20% on it vs 58% otherwise
  POOL WIDTH      -29 LP / 10 games     top 3: 20W-15L  ·  the other 6: 3W-13L

• What a champion is worth, in LP, per ten games on it. Not a win rate - the ladder. Your
  own measured LP (+23 / -17, read from your rank snapshots, not "about 20") against your
  own baseline. QUEUE Sett and BENCH Darius now sit on your profile with the number and
  the receipt attached, and the price appears right on the splash card for each champion.
• Compared against YOU, not against 50%. A 47% champion is not a leak for a player who
  wins 44% of everything else - it's their second-best pick, and every tool that flags it is
  wrong. Every comparison here is this champion vs your other games.
• POOL WIDTH: what your whole pool costs you. Your top 3 against everything else you
  queued, priced per ten of your games - the actual cost of ranked tourism, measured on you
  instead of quoted from a study. When your top 3 hold 80%+ of your games it says so and
  tells you there's nothing to cut.
• Your main is never benched. A main on an awful run is variance, and the card says
  exactly that: rough patch (6W-20L) - variance, not the pick. Keep queueing it. The old
  coach would tell you to ease off the champion you're best at, which is the worst advice
  the app was capable of giving.
• It reaches you in champ select. Hover a champion your own history has priced and the
  panel says so while you can still change your mind - ⚠ Darius: -93 LP / 10 on it (2W-12L) - in red, on its own line. Your own receipt outranks a study, so the 1M-game
  mastery warning is now the fallback, for champions your history can't speak about yet.
• One brain. The recommender's veto IS this board's bench. The page that tells you to
  bench a champion and the recommender that quietly stops suggesting it were two separate
  pieces of math that could disagree about the same champion. They can't any more.

And the part that matters most: it corrects for the fact that it is looking at all of your
champions at once. This is the failure mode of every "your best champion" stat ever put in
front of a League player, and it is not a rounding error - the guard suite measures it. On
random histories where every champion is a true coin flip, testing each one at the ordinary
bar declares a "proven" best or worst champion in 49% of them. Half. So the bar every
champion has to clear is divided among the champions being tested, in both directions - one
eligible champion needs z>=1.63, three need z>=2.11, eight need z>=2.48 - which takes that
same measured false-positive rate down to 7%. Underneath that bar a row still shows you
its numbers, labelled as what it is: a lean, never a price. A champion needs 6 games before
it can say anything at all, the whole board needs 15, and a one-trick gets told honestly that
there's nothing of yours to compare against. The pool-width test is a single question asked
in advance, so it correctly takes no such correction - and it refuses to price a "top 3" when
your games are spread evenly enough that which three is an alphabetical accident.

Tested: 15 guard groups plus a 1,200-pool fuzz that re-scores every random history
twice - once corrected, once not - and fails if the correction isn't measurably doing its
job, so this can never quietly rot into the stat it was built to replace. Also asserted, on
every shape of history that can exist: no priced claim ever escapes below its bar or its
sample; an earner is always positive and a bench row always negative; a main is never benched;
the same champion can never be both the queue and the bench; a champion price is always
quoted "per 10 on it" and a width price always "per 10 games", because mixing those two units
would be a lie; the price is always strictly below the flattering arithmetic on the raw gap;
and the profile page and the recommender are checked against each other champion by champion,
in both directions. A pool of junk - null rows, a champion claiming more wins than games, a
non-numeric average, an empty list - can't crash the board or fake up a sample. Every render
state was drawn and looked at.

Two bugs found while building it, both fixed. The champ-select note is drawn on one
unwrapped line and had only ~148px of room, so anything longer than about 28 characters was
being silently clipped mid-word - the team-scout roster line and the climb warning have
both been losing their tails there for releases. It now gets its own full-width line, and
truncates with an ellipsis when it has to, so a cut-off line reads as "go look at the page"
instead of as a rendering bug. And that line was drawn in green regardless of content -
including for the sub-12k mastery warning. A caution in the "all good" color is a caution
nobody reads; it now takes its tone from what it's actually saying.

And a tripwire for the release process itself: the self-test now asserts that every
core/ and ui/ module is in the frozen build's hidden-import list. Five modules were only
being included by luck. A module the packager misses ships an installer that crashes the
moment you use the feature - which is the worst possible way to ship a release named after it.

Smiteless v0.9.70

• THE ONE FIX: your leaks, priced in LP, and the single one worth fixing

New feature, and it answers the question the app has been dodging since the leak ledger
existed: you have five habits being graded every game - which ONE is actually costing you
the ladder, and what is fixing it worth?

Smiteless has spent five releases building a live guard for every leak your profile can
find. That work is done. What it never did was rank them. Your profile would hand you
three PATTERN bullets about the game you just played and leave you to guess which mattered.
Five things to fix is the same as no things to fix.

So the ledger it has been quietly writing all along - which habits were gradable each game,
which ones fired, and whether you won - is now read the other way round. Both sides of
every split are in there, so each leak can be priced:

RE-ENTRY      -41 LP / 10       with it: 3W-7L  ·  without: 9W-5L
FIRST TEN      7 of 24          with it: 4W-3L  ·  without: 8W-9L
BLEEDING       clean
CLOSING        clean
VISION         clean

• A number nobody has ever put in front of you: what a habit costs you in LP. Not a
  win-rate curiosity - the actual ladder. It fires in 4 of every 10 of your games; in those
  games you win 25 points less; a flipped game is worth what a win and a loss are worth to
  you. That is LP per ten games, and it is arithmetic on your own history, not a study.
• Your LP, measured, not assumed. Smiteless already snapshots your rank every time your
  profile builds, so it knows what a win and a loss are actually worth at your MMR - it
  reads +22 / -17, not "about 20". Only if it can't measure yours does it assume, and
  then it says so on the card.
• One fix. Not five. The board names exactly one habit, writes it as something you can
  actually hold yourself to for a single game - "after you respawn, 90 seconds of farm
  before you look for a fight" - and tells you which in-game guard is watching it for you.
  The other four sit underneath it, ranked, so you can see it was chosen and not asserted.
• And it meets you in the lobby. The QUEUE CALL card now carries a THIS GAME line:
  one habit, in the four seconds before you press Find Match, which is the only moment
  you're in a position to decide anything. It never appears under a STOP or a WAIT -
  handing you homework in the same breath as "log off with the LP" is how a stop rule gets
  ignored.
• Watch a leak close. Every row carries a form strip - the last six games that habit
  could have happened in - and flags improving or slipping once there's enough on both
  sides to mean it. A leak you are actually fixing visibly empties out.
• It fills itself in. The ledger used to learn exactly one game per profile open -
  forty opens before it could say anything. It now grades the games already on your profile
  page, so the board is live almost immediately, and it remembers 120 games instead of 60.

It refuses to overstate itself, which is the whole point. A leak is only priced in LP
once its split beats the same significance test the QUEUE CALL uses before it will tell you
to stop playing - and even then the split is quoted shrunk toward your own baseline, so a
3-games-versus-5 fluke can't become a 200-LP headline while a 40-game pattern is barely
touched. Below that bar the row still shows you its numbers, labelled as what it is: a lean,
not a price. A habit that fires in nearly every game can still be picked - on the count,
which needs no test at all - and then it quotes the count and never an LP figure. Under ten
graded games the board doesn't guess; it tells you how many more it needs. And it says out
loud, on the card, that this is a correlation from your own games.

Tested: 14 guard groups, including a 3,000-ledger fuzz that asserts - on every
random history it can build - that no LP price is ever attached to a claim that didn't clear
the test, that a priced row always carries a real price (a -0 LP row is a broken claim, so
it's demoted to a lean), that the headline, the board and the lobby line can never name
different leaks, and that no shape of garbage in the ledger can crash the page. Also
checked: a proven leak always outranks a merely frequent one; unfinished games are holes,
never evidence, and can't move a price; the ledger merge is idempotent and stays in time
order even when the backfill arrives newest-first; the LP reader rejects an MMR reset as
"not a game"; and the leak catalogue is verified against both the review page's wording and
the five shipped in-game guards, so nothing on the board can ever point at a surface that
doesn't exist. Every render state was drawn and looked at.

Smiteless v0.9.69

• THE WARD CLOCK grows up: the deadline, your trinket, and the 75 gold you keep forgetting

A big upgrade to the guard that shipped last version, in the four places it was still
leaving something on the table. If you play jungle or support, this is the one to read.

v0.9.68 gave the two map-owning roles their first coach: the live vision war (your score
against the enemy in your own role), PIT before a fight you'd be taking blind, and
PINK for a control ward rotting in your bag. What it didn't do was tell you when the
vision had to be in, how to place it with what you were carrying, or how much of the game
you'd spent with no control ward at all. All four are in now.

• The deadline, not just the countdown. A ward that goes down as the drake spawns is
  decoration. So the PIT card now back-times the moment vision actually has to be in - off
  the app's own objective clock, the same one drawing your timer chips:

  PIT - drake in 68s · in by 9:23 and nothing of yours is alive        1:00
  

  Inside the last seconds it drops the deadline and just says now, because by then that's
  the only honest answer.

• It reads the trinket in your hand. The same pit, three different instructions,
  because the item you're holding changes what "go ward it" even means:

  go PAST the pit - their bot jungle entrance, so you see them walk in
  and the fight starts on your terms · sweep it before you place -
  take theirs, don't just add yours
  

  Carrying a sweeper and it tells you to take theirs first: a swept pit beats a pit with
  one more ward in it. Carrying a farsight and it tells you to place from range and
  never to sweep, because you can't. A plain yellow trinket adds nothing, so it says nothing.

• The control ward in your bag is now a ledger, not a snapshot. A control ward leaving
  your inventory is the only evidence anywhere that you actually placed one - so the app
  now counts both halves, and the cards carry it:

  1 of 2 placed · control ward on you 42% of the game
  

  That second number has never existed. Not on the scoreboard, not in a post-game, not
  in any other tool: the share of the game you had 75 gold of map control on you at all.
  Most vision leaks are not bad ward placement. They are never having one in the bag - and
  now that's a number you can watch instead of a habit you can't see.

• It asks for the buy at the only moment you can act on it. Standing in the shop on a
  recall, with no control ward and the gold for one, the row leads with +75g control ward - and then drops it again. It never asks while you're already carrying one, and
  never when you can't afford it.

• It says it out loud. "Ward it." - at most three times a game, on the same contract
  BLEED's callout uses. This is the one guard whose entire subject is somewhere your eyes
  are not, which is exactly why it shouldn't need your eyes to land.

Everything that made the guard safe last version is untouched: it stays completely silent
until the live feed has proven it reports a vision score, so it can never accuse a support
who has warded all game of being dark; dark time you spent on the grey screen is never
billed to you; it defers to BLEED, RE-ENTRY and THE CLOSER for the card and stands down to
its quiet row the instant the tempo engine calls a fight; and it never says a word to a
laner, because your profile has never graded one on vision.

Tested: the guard's permanent suite grew to 24 fixtures and six full simulated games,
and every one of the four new behaviours is asserted on the text it prints, not just on which
verdict fired: the deadline must equal the objective clock's own lead (and must disappear
once the fight has started), a farsight must never be told to sweep, a sweeper must always be
told to take theirs first, the percentage must never print before there's a minute of game to
measure and must never escape 0-100, and the buy prompt must never reach somebody already
carrying a ward or unable to afford one. The purchase ledger is checked against the truth over
a 25-minute simulated game (two bought, one placed, carried 40% of the game - the number has
to agree). The inventory reads survive a null item row, a non-numeric item id and a missing
items list. Everything from v0.9.68 still holds: the arming tripwire, the dead-time freeze
proven neither to reset nor to accrue across four death windows, ambiguous lobbies returning
no counterpart rather than a wrong one, and total silence for all three laners. Every state
was rendered through the real widget path and looked at.

Smiteless v0.9.68

• THE WARD CLOCK: the vision war, live - and the last leak in the ledger

New feature, and it finishes the set. Every tag your profile can give you now has a
surface that fires while the mistake is still preventable.

no vision setup has sat in your ledger with nothing in the game to answer it since the
ledger existed. It was the last one. BLEED watches your health bar, RE-ENTRY the 90s after
a death, THE CLOSER the closeout, THE GOLD CLOCK the first ten minutes of farm - and the
gold clock is deliberately silent for jungle and support, because camps aren't on the lane
schedule and a support's CS was never the story. So the two roles whose entire job is
vision were the two roles with nothing to read. This one is theirs.

• It is a measurement, not a guess. :2999 reports a vision score for all ten
  players, every tick, unfiltered by fog - and that number only ever goes up while a
  ward of yours is alive. So a score that hasn't moved in 1:40 isn't an opinion. It is a
  fact that nothing of yours is on the map, and it needs no modelling constant at all.

• The head-to-head nobody has ever shown you. One quiet row, all game: you against the
  enemy in your own role. Same job, same minutes, same units - so it's an exact comparison,
  not a benchmark:

  WARD   14.2 v 21.6 · 0.9/min, bar 1.2 · 1 pink
  

  That is the only live scoreboard in League for the thing a support is actually doing, and
  the bar is your own profile's (1.2/min support, 0.55 jungle - the same number the
  no vision setup tag grades you on, read from one place so the two can never disagree).

• PIT - the fight you were about to take blind. In the ~75 seconds before a drake,
  grubs, herald or baron, if nothing of yours is alive, it takes the card:

  PIT - drake in 40s and nothing of yours is alive          1:00
  go PAST the pit - their bot jungle entrance, so you see them
  walk in and the fight starts on your terms
  

  It speaks on a shorter fuse there than anywhere else on purpose: "ward the pit before
  the drake" cannot be wrong, so a false alarm costs nothing and staying quiet costs the
  objective. And it tells you where - deep past the pit when you're ahead, your own tri
  and the pit mouth when you're behind, because those are opposite instructions and the
  right one depends on the game you're in.

• PINK - 75 gold of map you already paid for. It reads your actual inventory. A control
  ward you bought and have carried for two minutes gets one card, once, and then never
  again for that ward.

• It cannot cry wolf. The whole surface stays asleep until the live feed has proven
  it reports a vision score at all. If Riot ever drops the field, or a lobby doesn't send
  it, everything here degrades to silence instead of telling a support who has warded all
  game that he's blind. That tripwire is checked every single tick.

• It bills you once. Dark time you spent on the grey screen is never counted against
  you - that's RE-ENTRY's and BLEED's subject, and charging you twice for one death is how
  a coach gets switched off. It also hands the slot straight back the moment the tempo
  engine calls an actual fight, and it never outranks BLEED, RE-ENTRY or THE CLOSER.

• Jungle and support only. It stays silent for every laner rather than invent a number,
  for the same reason your profile has never graded a laner on vision.

• On by default: Settings -> Ward clock (the vision war, jg / sup). The widget legend
  has a new WARD CLOCK section. Included in MAX ELO.

Tested: 419,160 assertions before this shipped, plus the permanent guards in
selftest.py. Every verdict branch is driven across a full grid of role x clock x vision
score x dark time x control wards x objective state x tempo phase (and both armed states);
the pit window is checked against the app's real objective clock at every 7 seconds of a
40-minute game rather than hand-written fixtures; the counterpart matcher is asserted to
return nothing rather than guess whenever the lobby is ambiguous; and the guard is
driven second-by-second through whole simulated games - a support who wards constantly
(never accused once), one who stops at 6:40 (caught on the exact second the clock earns
it), one who never wards at all, a jungler on his own bar, four different death windows
(the dark clock must freeze, not reset and not accrue), a bought-carried-placed control
ward, all three laners (total silence), and a game where the feed reports no vision score
at all (total silence). A 20-shape malformed-payload sweep x every tempo/objective/win-read
combination must never raise. Two real defects were found and fixed by that sweep before
release, and every frame the guard produces was rendered through the real widget path and
inspected.

Smiteless v0.9.67

• THE GOLD CLOCK: your lane, counted against the minions that actually spawned

New feature, and it covers the biggest leak in your ledger that had nothing in the game
to answer it - the one almost every game you play can earn.

Your profile has tagged games weak first-ten economy since the behaviour ledger existed:
finish minute 10 under 55 CS and under 3100 gold and it goes in the file. Four of the five
roles can earn it. Nothing in the app has ever said a word about it while you could still
fix it - BLEED watches your health bar, RE-ENTRY the 90s after a death, THE CLOSER the
closeout. The thing you spend the entire lane phase actually doing was the blind spot.

• It does the arithmetic instead of guessing. Every CS overlay ever built shows you
  CS/min against a flat benchmark, and a flat benchmark is a bad coach - it doesn't know
  that at 4:32 only nine waves have spawned. Minions are a schedule: one wave leaves at
  1:05 and one every 30s after, 3 melee + 3 casters, every third wave carrying a cannon.
  So the denominator isn't a benchmark, it's the minions that have walked into your lane:

  GOLD   41 of 74 · 55% · on track for 63, bar 55
  

  Melee are 21g, casters 14g, cannons 60g, and every one of those is flat until 15:00 -
  which is why this whole window can be priced exactly rather than modelled.

• It back-times the deadline. 55 by 10:00 is your own tag's bar. Subtract what you have,
  divide by what's still coming, and you get the only sentence that actually helps:

  MISS - that wave went by · on track for 42 at 10:00, bar 55
  you need 25 of the next 32 minions (78%) · shove, then cross to a camp -
  a wave you walk away from is 105g on the floor
  

  And when the answer is no it says so - "49 short of 55 with 32 minions left" - and
  switches the advice to plates and objectives instead. A lane you cannot farm your way out
  of needs a different plan, and four more minutes of "farm harder" is four minutes wasted.

• The cannon minion has a clock and now you get it. 60 gold, the biggest single object
  in lane phase, and the one you most often give away because you were walking back from a
  roam. It warns you seconds before it lands - and only while you're behind, because a
  reminder that fires every 90 seconds regardless is a reminder you stop reading.

• Roaming never reads as farming badly. 30 CS and three kills is not a weak first ten,
  and your own tag agrees (it needs the gold bar missed too). Kills and assists are priced
  back into CS at the app's own per-CS rate - so the row shows 30+44 of 82 · 90% and
  stays green. The number it converts at is derived from the live gold model rather than
  typed twice, so the two can never drift apart.

• It bills you once. A wave you lost on the grey screen is never counted against you -
  that's RE-ENTRY's and BLEED's subject, and charging you twice for one death is how a coach
  gets switched off.

• It is a row, not a nag. For the whole ten minutes it's ONE quiet line you can glance
  at. It only takes the card at the moment a wave actually went by, or just before a cannon,
  and it hands the slot straight back. It never outranks BLEED (something that can kill you
  beats 105 gold) and never talks over a live objective call - the grub fight is the
  reason you left the wave.

• Top / mid / ADC. It stays silent for jungle and support rather than invent a number:
  camps aren't on the lane schedule and a support's CS was never the story.

• On by default: Settings -> Gold clock (farm pace, first 10 min). The widget legend has
  a new GOLD CLOCK section. Included in MAX ELO.

Tested: 10,721 assertions before this shipped. The wave schedule is verified spawn-by-
spawn and arrival-by-arrival for all three lanes across 30 waves (mid meets at 1:30, the
side lanes at 1:38, and a wave is never counted a tick before it lands); the cannon clock is
checked at every second of a 15-minute game for going negative, naming a non-cannon wave, or
pointing at a wave that already arrived; the gold composition is re-derived independently at
every sampled second; the per-CS conversion is asserted equal to the live gold model's own;
every verdict branch has a fixture; and the guard is driven through four full simulated games

• a clean farmer (never accused once), a player who quits farming at 5:00 (caught on the
  first wave boundary after he crosses the bar, not before), a player dead across three waves
  (billed for none of them), and an ADC - plus an eleven-payload malformed-data sweep that must
  never crash the widget. Every frame the guard produced was rendered through the real widget
  path and inspected.

Smiteless v0.9.66

• THE CLOSER: the games you were winning, and lost

New feature, and it only ever shows up in games you are already winning.

Smiteless has had a guard for every phase of a game that goes badly. It has never had one
for the phase that costs the most LP: ahead, past 20 minutes, not ended yet. That's the
game you already paid for. Every one of those you lose is LP you earned and handed back.

Your own ledger has been saying so the whole time. The profile tags a game threw_ahead
when one of your deaths lands after 25:00 with your team 2k+ up. It was the last tag in the
file with nothing in the game to answer it - RE-ENTRY owns the 90 seconds after a death,
BLEED owns the first fourteen minutes, the tempo engine owns the objective windows, and the
closeout was empty. Not any more.

• It reads the actual map. Every turret and inhibitor kill is in the live event feed and
  Smiteless was throwing all of it away. The CLOSER now keeps a live structure map of both
  bases and tells you the shortest path to their nexus:

  END IT - mid inhib open 3:34 - 2 of them dead
  nexus turrets as five - baron is a detour, the inhib clock isn't
  

  An open inhibitor is a five-minute clock and most games spend it taking baron instead.
  When one is down it says so, with the seconds left on it. When one turret stands in front
  of one, it tells you that turret is the game - not the next skirmish, not the drake.

• It tells you what you've GIVEN BACK. Nothing else in the app - and nothing else you
  can see in a game - tracks the trajectory of a lead. The CLOSER remembers your peak and
  shows the erosion: +4.5k - gave back 2.1k of 6.6k. A lead that has been quietly bleeding
  for four minutes is a game being thrown in slow motion, and you almost never notice from
  inside it. Once you've given back 1.5k it also tightens its own bar: an even fight is
  no longer good enough, because you don't need a fight, you need the nexus.

• It prices your death in seconds, live. Not "don't die" - the number:

  HOLD - you're +4.5k and down 2 bodies
  dying here costs 51s - baron is up inside that
  

  That's your real death timer at your level and this clock, checked against the live baron
  timer. Fifty-one seconds is what they buy with your one bad step, and now you can see it
  before you take it.

• It stays out of your way. Behind or even, it says nothing at all - a closeout coach
  talking during a losing game is worse than no coach. Ahead with nothing burning, it's one
  quiet line with your lead and your give-back. It never argues with the tempo card either:
  if the fight math says you win the fight, it will not tell you to hold.

• Late-game advice, finally. The "do this instead" lines are written for minute 28, not
  minute 8 - "stay behind your frontline - you never walk in first from ahead", "ward
  their jungle entrances - no solo roams" - instead of the lane-phase wave advice the other
  guards give.

• On by default: Settings -> Closer (win-conversion, from 20:00). The widget legend has a
  new CLOSER section explaining every verdict.

Tested: 60 assertions covering the structure parser (including a turret name Riot renames,
a replayed event, the fountain shrine, and the mid lane running past three turrets into the
nexus pair), the inhibitor clock from both the five-minute rule and the respawn event, all
twelve verdict branches, a full simulated game timeline, and a malformed-payload sweep that
must never crash the widget. The self-test now guards the CLOSER and the BLEED guard,
which had never had a check of its own.

Smiteless v0.9.65

Runes now adapt to the enemy comp - Talon takes Conqueror into a tank comp instead of always importing Electrocute. Only fires on an unambiguous comp, and cites op.gg's own numbers.

Smiteless v0.9.64

The recommender now reads your own results: vetoes champs you're proven bad on, and promotes ones you're good on but haven't played (the boredom fix). Ghost now sits on your Flash key.

Smiteless v0.9.63

Auto-mute now waits for a quiet moment before typing and aborts instantly if you touch the keyboard or mouse - your input can no longer collide with the command.